Add `min` and `max` rev specializations?

From this stackoverflow post it looks like we can phrase max(a,b) into

$$
max(a,b) = \frac{a + b + |a-b|}{2}
$$

and similarly for min

$$
min(a,b) = \frac{a + b - |a-b|}{2}
$$

So except for the case when $a=b$ this function is differentiable. For max we have

$$
\begin{align}
\frac{\partial max}{a} = \frac{1}{2} + \frac{sign(a-b)}{2}
\frac{\partial max}{a} = \frac{1}{2} + \frac{sign(a-b)}{2}
\end{align}
$$

so if a>b the gradient is 1 and if a<b the gradient is 1/2. For min we have

$$
\begin{align}
\frac{\partial min}{a} = \frac{1}{2} - \frac{sign(a-b)}{2}
\frac{\partial min}{a} = \frac{1}{2} - \frac{sign(a-b)}{2}
\end{align}
$$

where if $a>b$ the gradient is 0 and if $a<b$ the gradient is 1/2

From this it seems like pytorch does something similar to the above. I think this would be nice to implement. We can also write min(vector) for these as well.

Read the linked Stack Exchange and PyTorch discussions first, then inspect the existing min/max APIs and reverse-mode specialization patterns in the library. Clarify the intended scalar and vector behavior, including equal arguments and the proposed gradients; done means the agreed specializations are implemented with tests covering those cases.