It would be great if they existed, even if only for the symmetric group. One possible application would be to algebraic combinatorics: Mark Haiman has collected data suggesting that the ring $R/R^{S_n}_+$ has dimension $2^n (n+1)^{n-2}$, where
$$R=\mathbb{Q}[x_1,\dots,x_n,y_1,\dots,y_n,z_1,\dots,z_n],$$ the symmetric group acts by simultaneously permuting the three sets of variables, and the notation $R^{S_n}_+$ means the ideal generated by positive degree symmetric polynomials.
Iain Gordon showed that it is possible to use the representation theory of the rational Cherednik algebra (the rational object in the rational/trigonometric/elliptic trichotomy whose elliptic object is the DAHA) to establish the correct lower bound on this dimension in the case of two sets of variables (this dimension turns out to be $(n+1)^{n-1}$, a theorem proved by Haiman using a suggestion of Procesi and the geometry of the Hilbert scheme of points in the plane). So one might hope to use the representation theory of a TAHA to prove the correct lower bound, at least.
The reason the double affine Hecke algebra exists at all is a little subtle, and has to do with @Theo Johnson-Freyd's comments to the question: the affine Hecke algebra has two realizations. First, the affine Hecke algebra is an affinization of a finite Hecke algebra; second, it is the Hecke algebra associated to the affine Weyl group (or, if you prefer, for a certain specialization it is the Hecke algebra corresponding to an Iwahori subgroup of a p-adic group). Starting with the second presentation, one affinizes again to obtain the DAHA. The point here is that
$$\{\text{affine Hecke algebras} \}=\{\text{Hecke algebras of affine groups} \}$$
and we know how to affinize the Hecke algebras on the RHS.
But so far the DAHA has no second realization as the Hecke algebra of something that can be affinized again. Perhaps recent work of Kazhdan and his collaborators could help here, but I have not read these papers carefully enough to know.
There is a second approach that is somewhat more geometrical. The rational Cherednik algebra is a deformation of the algebra $\mathbb{Q}[x_1,\dots,x_n,y_1,\dots,y_n] \rtimes S_n$, so one might look for nice deformations of the analogous objects in three sets of variables. Perhaps experts in Hochschild cohomology have done calculations suggesting where to look?
The specific issue of what automorphic forms on bigger groups than $GL(2)$ over $\mathbb Q$ (for example) may tell us about automorphic forms (and L-functions) for $GL(2,\mathbb Q)$ or $GL(1,k)$ for number fields $k$ does have at least a few good answers. First, about 1960 and a little before, Klingen's proof that zeta functions of totally real number fields $k$ (and L-functions of totally even characters on such fields) have good special values at positive even integers used the idea of pulling back holomorphic Hilbert modular Eisenstein series to elliptic modular forms. (I heard G. Shimura lecture on this c. 1975, and it was quite striking.)
Another example: already in the 1960s, J.-P. Serre and others saw that holomorphic-ness of symmetric-power L-functions for $GL(2)$ holomorphic modular forms would prove Ramanujan-conjecture-type results. How to prove that holomorphy? By finding an integral representation of such L-functions, and using that. This has met with varying degrees of limited success, e.g., in papers of H. Kim and F. Shahidi.
The previous example was grounded in the general pattern of Langlands-Shahidi treatment of L-functions in terms of constant terms of cuspidal-data Eisenstein series on (necessarily larger) reductive groups. The specifica cases where various Levi-Malcev components of parabolics were products of $GL2$'s or $SL2$'s produced several "higher" L-functions for $GL2$.
As variation on that, already in the Budapest conference in 1971, Piatetski-Shapiro observed that (what we often nowadays call) Gelfand pairs could produce Euler produces via integral representations (usually involving Eisenstein series). Various success-examples of this idea included work of PiatetskiShapiro-Rallis, Shimura, myself, M. Harris, S. Kudla, various collaborations among these people, and several others, beginning in the late 1970s. E.g., I was fortunate enough to stumble upon an integral representation for triple tensor product L-functions for $GL2$ in terms of an integral representation against Siegel Eisenstein series on $Sp(3)$ (or $Sp(6)$, if one prefers). M. Harris and S. Kudla found another such integral representation that covered special value results in the "other range" (in terms of P. Deligne's conjectures).
In yet other terms, Jacquet-Lapid-Rogawski (and several others) have demonstrated that a variety of L-functions appear as periods of Eisenstein series on "larger" reductive groups. (One novelty is using relative trace formulas to exhibit Euler products when the simpler "Gelfand pair" idea is not quite sufficient.)
Best Answer
In addition to many excellent answers posted so far, I would like to explain another way in which the relations of the "graded" or degenerate affine Hecke algebras arise in representation theory, which I find most helpful in understanding them (and is how I would recover them for myself if stranded on a desert island). This introduction has the advantage that it could be explained to an undergraduate who knew what a surface and a fundamental group was. All of this is in Cherednik's book in one form or another.
In this point of view, the "full" DAHA and AHA are the primary objects, and one recovers the degenerate versions by a process similar to the degeneration from $U_q(\mathfrak{g})$ to $U(\mathfrak{g})$. Let me just discuss $A_n$, although an analogous construction works for any Weyl group (and parts of the discussion for any complex reflection or symplectic reflection group).
The Artin braid group $B_n$ is $\pi_1(C_n(\mathbb{C}))$, the fundamental group of configurations of $n$ points in $\mathbb{C}$. It's easy to identify this with the braid group associated to the root system of type $A_{n-1}$, since the reflection hyperplanes are precisely what impose the distinctness of the points in the configuration. Denote the usual generators of $B_n$ by $T_i$.
One may instead consider the "double affine braid group" $DB_n$, $\pi_1(C_n(E))$, which is the configuration space of $n$ points on an elliptic curve (or what matters topologically is that it's a $S^1\times S^1$). One finds generators $B_n$ corresponding to loops which are contained in a contractible ball, and new generators $X_1,\ldots, X_n, Y_1,\ldots, Y_n$ corresponding to taking the $i$th point of the configuration and running it around the the inside our outside ring of the torus. One computes relations that the $X_i$ commute, the $Y_i$ commute, and $T_iX_iT_i=X_{i+1}$, $T_iY_{i+1}T_i=Y_i$, and (Y_2,X_1):=Y_2X_1Y_2^{-1}X_1^{-1}=T_1^{-2}$.
Now, how is this related to the algebras in your question? Let $\tilde{A}$ denote the group algebra of $DB_n$ with coefficients in $\mathbb{C}[[\hbar]]$, completed in the $\hbar$-adic topology. Let $A$ denote the quotient by the additional relations
$(T_i-q)(T_i+q^{-1})$, where $q:=e^{\hbar/2}$.
Now suppose that $V$ is some representation of $A$ such that $Y_i$ acts as $1$ modulo $\hbar$. In this case, it makes sense to define $s_i$ and $y_i$ in $A$ by the relations:
$Y_i=e^{\hbar y_i}$, $T_j:=s_je^{k\hbar s_j}$.
Where I'm evaluating in the representation $V$ so that the first equation makes sense, but I'm not writing that in explicitly.
One now would like to check what relations are imposed on the generators $X_i$, $y_i$ and $s_j$ so defined. Let us just check what relations we get by considering the relations of $A$ up to first non-trivial order in $\hbar$. We find:
$(T_i-q)(T_i+q^{-1})=0 \Rightarrow s_i^2=1$.
Braid relations for $T_i \Rightarrow$ braid relations for $s_i$.
$X_i$'s commute (as before). $\tilde y_i:=y_i + \sum_{i `< j} s_{ij}$.
$T_iX_iT_i=X_{i+1}\Rightarrow s_iX_is_i=X_{i+1}$
$T_iY_{i+1}T_i=Y_i\Rightarrow s_iy_is_i=y_{i+1}+s_i$
$(Y_2,X_1):=Y_2X_1Y_2^{-1}X_1^{-1}=T_1^{-2}\Rightarrow [y_2,X_1]=ks_1X_1$
which are (one form of) the relations of the trigonometric Cherednik algebra.
Further writing $X_i=e^{\hbar x_i}$, one recovers the so called rational Cherednik algebra.
So there is a hierarchy of degenerations. The top and bottom of the hierarchy are essentially symmetric in the variables $X$ and $Y$, in the precise sense that there is a "Fourier transform" automorphism swapping the variables, in both cases. Note that at the top of the hierarchy, the Fourier transform is just the order four automorphism of the elliptic curve which is the matrix $((0,1),(-1,0))$ in $PSL_2(Z)$, the mapping class group of the torus, and which can be seen in various elementary ways. At the bottom of the hierarchy, the Fourier transform is just swapping $x_i$ and $y_i$ and is related to Fourier transform of differential operators on an abelian group. In the middle, there isn't really a Fourier transform, because the symmetry was broken by degenerating the $Y_i$, but leaving the $X_i$ unscathed.
I'm not an expert on any of these things, and I can't say I've worked through the topological construction for all the cases, but I think, to answer Stephen's question, we can simply choose any lattice in $\mathbb{C}$, of rank $2$ (say $\Lambda=\mathbb{Z}1 \oplus \mathbb{Z}\mathbf{i})$, and consider $\pi_1(\mathfrak{h}_{reg}/W\ltimes\Lambda^r)$, where $r$ is the rank of $\mathfrak{h}$. Then one quotients by the relations that loops around singular points of this quotient have order 2. For instance, let me explain a nice thing that happens for BC_n type.
BC_n non-affine braid group means we configurations of 2n distinct points in $\mathbb{C}\backslash\{0}$, such that $x$ is in the configuration if and only if $-x$ is in the configuration. We don't allow zero because that we want pairs of matched points. It's not hard to see that $\pi_1$ of that configuration space is the braid group of type $BC_n$, because you can choose the repesentative of each pair lying in the upper half plane, and you get the upper half plane, except that the $r$ and $-r$ are identified for all real points, and zero is excluded. This is a punctured plane.
Now the prescription above should lead you to consider $2n$ distinct points on the elliptic curve (or rather $S^1\times S^1$; I only say elliptic curve because people sometimes mean $\mathbb{C}^\times$ by torus...). However, now you have more points to remove. Not only zero, but all half integer points would correspond to a place where $x$ and $-x$ collide. So you get the usual picture of the torus from a first course in topology, except with half-integer points removed. A very fun exercise is to work out that you can again choose the representative of each pair which lies in, say, the lower left corner of the torus, where we cut the torus in half along the diagonal from upper left to lower right. But now again you have to identify some edges, and you get .... $\mathbb{CP}^1$ with four punctures (corresponding to the four half integer points of your real torus.) Now you get to choose five parameters: one parameter for non-affine hyper planes, meaning the $T_i$, and then one for each of the new poles you've introduced. Imposing the Hecke relations on all those you get the double affine Hecke algebra of type $BC_n$ (sometimes called $C^\vee C_n$ for historical reasons) with those five parameters. Notice that this process leads you to a very different presentation of the DAHA, where you de-emphasize the lattices, and emphasize the loops around the singularities. I understand that S. Sahi gave a similar presentation for all DAHA's coming from root systems, but with probably deeper motivations than these drawings on surfaces. I'm not sure of the reference where Sahi did this.
At least in this example, one again sees the Fourier transform as just the obvious move on the real torus.