[llvm] [AMDGPU] Use 32-bit SGPR to save/restore of SCC (PR #68367)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm requested changes to this pull request.

The pre-existing code here is simply not sensible. It doesn't make sense to unconditionally replace SCC copies with a similar looking vector operation. For the purposes of the local change, the minimal fix is to just directly emit the s_cselect_b32. We should also untangle the current handling, because this really should just emit a copy 

https://github.com/llvm/llvm-project/pull/68367