[llvm] [AMDGPU] Fix GCNRewritePartialRegUses pass: vector regclass is selected instead of scalar. (PR #69957)

[Valery Pykhtin via llvm-commits]

vpykhtin wrote:

@arsenm let's submit this and I create another PR to cleanup unnecessary register class deduction, since _TRI->getSubRegisterClass_ works correctly now.

https://github.com/llvm/llvm-project/pull/69957