[llvm] [AMDGPU] Switch V_CNDMASK operands to shrink it into VOP2 (PR #135162)

[Mirko Brkušanin via llvm-commits]

https://github.com/mbrkusanin commented:

- Need rebase
- Please add tests for a negative case, case with source modifiers and when shouldSwapCndOperands counter is positive.

https://github.com/llvm/llvm-project/pull/135162