[llvm] [AMDGPU] Do not fold clamp when only one max source has modifiers (PR #224517)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm approved this pull request.

I'd keep the commuted test if it doesn't already exist 

https://github.com/llvm/llvm-project/pull/224517