decomposition for torch.minimum, torch.maximum

[t-vi]

torch.maximum, torch.minimum can be decomposed using where (though we need to be careful about NaNs), we should do that...

cc @apaz-cli @tfogal

[mruberry]

What's interesting about both torch.minimum and torch.maximum is that they are primitives in PyTorch, and similar operators are primitives in JAX.

While these operations can be expressed using other operations, it may be easier for executors to make the primitives fast than a decomposition of the operations.

[t-vi]

Works for me as long as NVFuser fuses them (I initially wanted to file that as an issue but then switched to the decomposition).

[mruberry]

I added the nvfuser label. Would you like to retitle the issue something "Add nvFuser support for torch.minimum and torch.maximum"?