Fix torch.clamp in MPS to handle NaN correctly (#121381)
Fixes #120899
So this is interesting. There are methods that specifically propagate NaN instead of clamping to real numbers.
https://developer.apple.com/documentation/metalperformanceshadersgraph/mpsgraph/3857573-maximumwithnanpropagationwithpri
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121381
Approved by: https://github.com/malfet
