Better ReLU formulation

[rdeits]

As pointed out by @vtjeng, doing

y = @switch((x <= 0) => 0, (x >= 0) => x) 

results in a model whose convex relaxation isn't as tight as it should be (specifically, it misses the constraints that y >= 0 and y >= x). Can we do better?

I think the ideal result would be:

y >= x
y >= 0
y <= x + M(1 - z)
y <= Mz