[llvm] [AMDGPU] Refactor GFX11 VALU Mask Hazard Waitcnt Merging (PR #169213)

[Carl Ritson via llvm-commits]

perlfu wrote:

> > In turn this greatly decreasing numbers of waits inserted in sequences of V_CMP instructions writing SGPRs improving VALU pipeline performance.
> 
> Can you precommit the new tests to demonstrate this improvement?

Rebased on to pre-commit test showing the removal over unnecessary waits.

https://github.com/llvm/llvm-project/pull/169213