[llvm] [AMDGPU] Use s_cmovk_i32 instead of s_cselect_b32 when applicable (PR #135232)

[Ryan Buchner via llvm-commits]

bababuck wrote:

I updated this to the most recent `main`. In addition, I added a new test and split this MR into 3 commits:
1. Adding the new MIR test.
2. Adding the logic for `S_CSELECT` to `S_CMOVK`.
3. Adding the logic for hinting that the source and destination in `S_CSELECT` should be assigned the same virtual registers so that we can convert to `S_CMOVK` later.

I split up commits 2 and 3 to demonstrate the impact of the hinting change, namely in the new test I added.

@jayfoad @arsenm would it be possible for this to get assigned to the appropriate people.

https://github.com/llvm/llvm-project/pull/135232