[llvm] [AMDGPU][True16] added Pre-RA hint to improve copy elimination (PR #103366)

[Jay Foad via llvm-commits]

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

Given that 16-bit values can live in either VGPR_16 or VGPR_32, this approach seems reasonable to me.

Eventually I think we should aim to clean this up such that 16-bit values always live in VGPR_16.

https://github.com/llvm/llvm-project/pull/103366