[llvm] [AMDGPU] Gate rootn(x, +-2) -> sqrt/rsqrt fold on nsz/ninf (PR #200578)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> _rootn_fast keeps the unconditional fold

Note the __fast variants are supposed to have correct edge case handling, only reduced accuracy 

https://github.com/llvm/llvm-project/pull/200578