As far as I know, all known quantum field theories have the same very broad structure: one gives some finite list of data in order to specify a particular QFT, then one uses some formalism to algorithmically extract various physical observables from that specifying data. (I suppose a philosopher might say that all physical theories follow this same basic pattern.)
But how this general framework works out looks very different for different types of QFTs:
 A "standard," Lagrangianbased QFT is specified by a particular choice of Lagrangian density containing only renormalizable interaction terms. More concretely, the "input data" that determines a theory is a list of fields and a finite set of coupling constants between those fields. The observables that come out of this framework are primarily $n$point correlation functions of the various fields. (Although not necessarily all possible correlation functions – for example, for gauge theories, only correlation functions of gaugeinvariant quantities are physically observable. These correlation functions are not always the "final answer" – e.g. we might plug them into the LSZ formula to get scattering amplitudes instead. And sometimes we might want to answer questions that are not directly answered by correlation functions – e.g. the sign of the beta function, or whether a particular theory experiences a phase transition. But in principle, the correlation functions (directly or indirectly) determine all observable quantities.)
 A conformal field theory superficially looks very different. We often do not write down any Lagrangian for a CFT, especially when working within the framework of the conformal bootstrap. "The full set of data and consistency conditions associated with a CFT is not known in general," as discussed here. But we believe that, at least for scalar fields on flat spacetime, the primaryfield conformal weights $\Delta_i$ and the operator product expansion coefficients $f_{ijk}$ together form sufficient CFT data. (This CFT data must respect certain consistency constraints, like the crossing symmetry equation (which ensures the associativity of the OPE), and modular invariance in two dimensions. There may be other consistency constraints as well.) But just like in the Lagrangian case, the observables that come out of the framework are typically $n$point correlation functions.
 A topological field theory again looks very different. There are several different ways to formulate TQFTs: in terms of symmetric monoidal functors, or braided fusion categories, or modular tensor categories (that's seven different links). There are also Schwarztype TQFTs, which tend to come up in condensedmatter theory, and Wittentype TQFTs, which tend to come up in highenergy theory. The basic idea is that a TQFT is a map from a spacetime topology to some topologically invariant complex number. Depending on the specific formulation, the initial data that specifies a particular choice of TQFT can be a set of $f$ symbols, or $S$ and $T$ matrices, etc., and the "observables" that come out are various topological invariants. There is usually no useful notion of $n$point correlation functions.
Obviously, it is not at all clear how to combine these three very distinct notions into a single unifying conceptual framework. I have two related questions about the relationships between them:

When used without qualification, the term "quantum field theory" usually refers to the Lagrangianbased formalism. We often intuitively think of the other two as special cases of this one. For example, we often think of a CFT as an RG fixed point of some nonconformal QFT, but we rarely write down the CFT Lagrangian explicitly. In fact, many CFTs have no known Lagrangian description at all. It is even suspected that some CFTs, like $(2, 0)$ superconformal field theory in six dimensions, have no possible Lagrangian description, although there is apparently no proven nogo theorem.
Similarly, Schwarztype TQFTs are often thought of as Lagrangianbased theories whose Lagrangians do not depend on the spacetime metric. In fact, one could perhaps think of all Schwarztype TQFTs as also being CFTs, since any Schwarztype TQFT's Lagrangian density trivially transforms conformally (i.e. $g_{\mu \nu}(x) \to \Lambda(x) g_{\mu \nu}(x)$) under arbitrary diffeomorphisms, since the metric does not appear at all! (Although again, in practice CFTs and TQFTs look very different.) Have either of the reasonableseeming inclusions $\text{TQFT} \subset \text{CFT}$ or $\text{CFT} \subset \text{(Lagrangian QFT)}$ been proven or disproven? (This question is related to this one.) 
Do attempts to mathematically formalize QFT (e.g. the Wightman axioms, etc.) try to cover all these types of QFT's simultaneously? If not, is there any mathematical framework that unifies them? It is my understanding that certain topological QFT's have been made completely mathematically rigorous, but little progress has been made for Lagrangianbased QFTs. I'm not sure what the status of CFT is from the standpoint of mathematical rigor.
Best Answer
The three classes of QFTs you are referring to are distinguished by different symmetry assumptions (Poincare invariance, conformal invariance, and volumepreserving diffeomorphism invariance) and different background spacetimes (Minkowski, Riemann curve (or families of them), and arbitrary manifolds). Moreover, Wightman axioms only characterize the vacuum sector of a Poincareinvariant QFT.
Each set of assumptions leads to very different natural questions and constructions, hence different mathematical approaches. This explains why there is this diversity of approaches. In the light of this diversity, a uniform theory would be conceptually very shallow  too general to be restrictive and hence useful  and would break immediately into chapters distinguished by specific assumptions.