[llvm] [AMDGPU][NewInsertWaitcnts] Boilerplate for the new wait insertion pass (PR #226335)

[Vasileios Porpodas via llvm-commits]

vporpo wrote:

> I am curious to read the underlying motivation and plan of action.

Motivation:
- Move away from the score-only tracking data structure to a less lossy tracker that keeps track of the actual instructions (among others) per counter and maintains a timeline that you can inspect and reason about. This will avoid the use of ad-hoc tracking data structures, will simplify the overall code of the pass and make it more maintainable in the long run. @Pierre-vh will follow up with several PRs on this new tracker soon.
- Overall redesign and cleanup of the pass with good separation of transformations and phases.

We will be developing this pass alongside the current one, and will start migrating targets once it is ready. The old pass will be retired shortly after all targets have migrated. 

https://github.com/llvm/llvm-project/pull/226335