[llvm] [AMDGPU] Prefer vector i8s in PHI Nodes (PR #91016)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > Vectors of i32 should be the canonically preferred type for anything wide. Why would we want i8?
> 
> I think we should prefer vectors of i8 over scalar i8s for PHIs because vectors of i8s for PHIs will be coerced to vectors of i32 to workaround the legalization issues. So, in that case, vectorization is profitable

Description should be clarified, because "for the vectorizers" and "for codegen" are different questions 

https://github.com/llvm/llvm-project/pull/91016