[llvm] [AMDGPU] Optimize away v_readfirstlane_b32 on SGPR input (PR #151033)

[Josh Hutton via llvm-commits]

JoshHuttonCode wrote:

I chose to handle this in the location I did because that is where the `v_readfirstlane_b32` was being inserted, but SIFoldOperands does seem like it may be a better fit. I will investigate if this could be moved there.

https://github.com/llvm/llvm-project/pull/151033