[llvm] [WIP][AMDGPU] Improve the handling of `inreg` arguments (PR #133614)

[Matt Arsenault via llvm-commits]

arsenm wrote:

Vaguely related, but we should also support inreg on function returns so we can avoid using the ugly hack of integer typed return values with shader calling conventions. Still not as nice as a proper per-field entry on the struct 

https://github.com/llvm/llvm-project/pull/133614