DIGRESSION
This is a good example to showcase that densities should better be defined for the whole real line using indicator functions and not branches. Because, if one looks at the likelihood, one could, at least for a moment, say "hey, this likelihood will be maximized for the value from the sample that is positive and closest to zero -why not take this as the MLE"?
The density for one typical uniform in this case is
$$f \left( x, \theta_x \right) =\frac{1}{2\theta_x}\cdot \mathbf 1\{x_i \in [-\theta_x,\theta_x] \},\qquad \theta_x >0$$
Note that the interval is (and should be) closed, and that we define the parameter as positive because, defining it as belonging to the real line would a) include the value zero which would make the setup meaningless and b) add nothing to the case except heavy dead-burden notation.
The likelihood of an (ordered) sample of size $i=1,...,n_1$ from i.i.d such r.v.s is
$$L(\theta_x \mid \{x_1,...,x_{n_1}\}) = \frac{1}{2^{n_1}\theta_x^{n_1}}\cdot \prod_{i=1}^{n_1}\mathbf 1\{x_i \in [-\theta_x,\theta_x] \}$$
$$=\frac{1}{2^{n_1}\theta_x^{n_1}}\cdot \min_i\left\{\mathbf 1\{x_i \in [-\theta_x,\theta_x]\}\right\}$$
The existence of the indicator function tells us that if we select a $\hat \theta$ such that even one realized sample value $x_i$ will be outside $[-\hat \theta_x,\hat \theta_x]$, the likelihood will equal zero. Now, respecting this constraint, this likelihood is always higher for positive values of the parameter, and it has a singularity at zero so it is "maximized" (tends to plus infinity) as $\theta_x \rightarrow 0^+$. It is then the constraint to choose a $\theta_x$ such that all realized values of the sample are inside $[-\hat \theta_x,\hat \theta_x]$ that guides us to move away from zero the minimum possible (reducing the value of the likelihood as little as possibly permitted by the constraint), and this is the actual reason why we arrive at the estimator $\hat{\theta_x} = \max \{-X_1,X_{n_1} \}$ and the estimate $\hat{\theta_x} = \max \{-x_1,x_{n_1} \}$.
MAIN ISSUE
To arrive at the joint distribution of $V$ and $U$ as defined in the question, we need first to derive the distribution of the ML estimator. The cdf of $\hat \theta_x$, respecting also the relation $X_1 \le X_{n_1}$, is
$$F_{\theta_x}(\hat{\theta_x}) = P(-X_1 \le \hat{\theta_x}, X_{n_1}\le \hat{\theta_x}\mid X_1 \le X_{n_1}) = P(-\hat{\theta_x} \le X_1 \le X_{n_1}\le \hat{\theta_x})$$
Denoting the joint density of $(X_1, X_{n_1})$ by $f_{X_1X_{n_1}}(x_1,x_{n_1})$, to be derived shortly, the density of the MLE therefore is
$$f_{\theta_x}(\hat{\theta_x}) = \frac {d}{d\hat{\theta_x}}F_{\theta_x}(\hat{\theta_x}) = \frac {d}{d\hat{\theta_x}}\int_{-\hat \theta_x}^{\hat \theta_x}\int_{-\hat \theta_x}^{x_{n_1}}f_{X_1X_{n_1}}(x_1,x_{n_1})dx_1dx_{n_1}$$
Applying (carefully) Leibniz's rule we have
$$f_{\theta_x}(\hat{\theta_x}) = \int_{-\theta_x}^{\hat \theta_x}f_{X_1X_{n_1}}(x_1, \hat \theta_x)dx_1 -(-1)\cdot \int_{-\hat \theta_x}^{-\hat \theta_x}f_{X_1X_{n_1}}(x_1,-\hat \theta_x)dx_1 + \\ +\int_{-\hat \theta_x}^{\hat \theta_x}\left(\frac {d}{d\hat{\theta_x}}\int_{-\theta_x}^{x_{n_1}}f_{X_1X_{n_1}}(x_1,x_{n_1})dx_1\right)dx_{n_1}$$
$$= \int_{-\theta_x}^{\hat \theta_x}f_{X_1X_{n_1}}(x_1, \hat \theta_x)dx_1+0-(-1)\cdot \int_{-\hat \theta_x}^{\hat \theta_x}f_{X_1X_{n_1}}(-\hat \theta_x,x_{n_1})dx_{n_1}$$
$$\Rightarrow f_{\theta_x}(\hat{\theta_x}) =\int_{-\theta_x}^{\hat \theta_x}f_{X_1X_{n_1}}(x_1, \hat \theta_x)dx_1+\int_{-\hat \theta_x}^{\hat \theta_x}f_{X_1X_{n_1}}(-\hat \theta_x,x_{n_1})dx_{n_1}$$
The general expression for the joint distribution of two order statistics is
In our case this becomes
$$f_{X_1,X_{n_1}}(x_1,x_{n_1}) = \frac {n_1!}{(n_1-2)!}f_X(x_1)f_X(x_{n_1})\cdot\left[F_X(x_{n_1})-F_X(x_1)\right]^{n_1-2}$$
$$\Rightarrow f_{X_1,X_{n_1}}(x_1,x_{n_1})=n_1(n_1-1)\left(\frac 1{2\theta_x}\right)^2 \left[\frac {x_{n_1}+\theta_x}{2\theta_x} - \frac {x_1+\theta_x}{2\theta_x}\right]^{n_1-2}$$
$$\Rightarrow f_{X_1,X_{n_1}}(x_1,x_{n_1}) = n_1(n_1-1)\left(\frac 1{2\theta_x}\right)^{n_1}(x_{n_1}-x_1)^{n_1-2}$$
Plugging this into the expression for the density of the MLE and performing the simple integration we get
$$f_{\theta_x}(\hat{\theta_x})=n_1\left(\frac 1{2\theta_x}\right)^{n_1}\cdot \left[-(\hat \theta_x-x_1)^{n_1-1} + (x_{n_1}+\hat \theta_x)^{n_1-1}\right]_{-\hat \theta_x}^{\hat \theta_x}$$
$$=n_1\left(\frac 1{2\theta_x}\right)^{n_1}\cdot 2\cdot (2\hat \theta_x)^{n_1-1}$$
$$\Rightarrow f_{\theta_x}(\hat{\theta_x}) = \frac {n_1}{\theta_x^{n_1}}\hat \theta_x^{n_1-1}$$
By the way , this is a "extended-support" Beta distribution, Beta$(\alpha = n_1, \beta =1, min=0, max = \theta_x)$. For the $Y$ r.v.s the expression would be the same, using $\hat \theta_y,\, \theta_y,\, n_2$.
We turn now to the joint density $g(u,v)$ as defined in the question. $U$ and $V$ are defined as the extreme order statistics of just ...two random variables, $\hat \theta_x$ and $\hat \theta_y$. If we want to apply again the theorem above in order to derive the joint density of $U$ and $V$ under the null $H_0$ (which only makes $\theta_x = \theta_y=\theta$), we need the densities of the MLEs to be identical, and for this we need in addition that $n_1=n_2=n$ (and the mystery is solved). Under these assumptions, application of the theorem (the $n$ of the theorem is now set equal to $2$) we get
$$g(u,v) = 2f_{\theta}(u)f_{\theta}(v)$$
$$=2\frac {n}{\theta^{n}}u^{n-1}\frac {n}{\theta^{n}}v^{n-1} = 2n^2u^{n-1}v^{n-1}/\theta^{2n}$$
Question 1 is straightforward, here $\alpha$ is a vector of repetitions of the given value. (As answered by Max S.)
Question 2 is more interesting: The Dirichlet distribution has the following interpretation relevant in this context: When $\alpha$ is the observed vector of outcome-counts drawn from some (unknown) categorical distribution with outcome probabilities $\pi$, then $Dir(\alpha)(\pi)$ is the likelihood that $Cat(\pi)$ is the actual underlying distribution given you observed $\alpha$ as the counts. (This is basically the definition of a dual distribution.)
Now P(s,a) estimates the probability that a good player would play a in s, that is the parameters of his categorical distribution, which AlphaZero wants to learn. So $Dir(\alpha)$ would sample reasonable estimates for $pi=$P(s,a) if we observed a good player play moves $\alpha$-times.
But if some $\alpha_i=0$, then all $\pi\sim Dir(\alpha)$ have $\pi_i=0$, preventing exploration. By adding the noise they assume that they have observed every move being played some small number of times $\alpha$ (here chosen 0.3, 0.15, 0.03).
As for how they got the constants, my guess is that they assume to have observed ~10 random plays in every game: In chess, $Dir(0.3)$ assumes that you have seen each move played 0.3 times. Given that there are ~35 moves available according to Allis, the authors assume you have seen ~10 random moves in every node. In Go, if we assume ~270 legal moves on average (3/4 of 361 board positions), we see an equivalent to observing ~8 random moves. (I do not have the data for Shogi.)
Best Answer
All points in the triangle must satisfy the two constraints: between zero and one in each dimension ($0 \leq \theta \leq 1$) and all sum up to one ($\theta_0 + \theta_1 + \theta_2 = 1$).
The way I finally understood it is the following:
So (a) shows a 3-D space with $\theta_{1, 2, 3}$ as coordinates. They range only between 0 and 1.
In (b), a triangle is shown, this is our simplex.
(c) shows two example points that "lay" on the simplex which also fulfil the second criteria (sums up to one).
(d) shows another example point on the simplex, the same constraints hold
In (e), I tried to show a projection of the simplex to a 2-D triangle with all example points shown before.
Hope it makes more sense now :)