[llvm] [AMDGPU] Flush denormal results when constant folding amdgcn.rcp (PR #203566)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm commented:

We probably should stop constant folding this if we can't guarantee bit identical results. Which this still can't. This also would need to consider the input mode 

https://github.com/llvm/llvm-project/pull/203566