[clang] [llvm] [AMDGPU] Introduce asyncmark/wait intrinsics (PR #173259)

[Sameer Sahasrabuddhe via cfe-commits]

ssahasra wrote:

Pushed a new single commit:

- Squash and rebase everything.
- Use explicit "async" builtins/intrinsics instead of aux/cpol bits.
- Enforce a maximum count on marks.
- Uniformly use the name "mark" instead of "marker". Saves a syllable when speaking, and everyone seems to say "mark" anyway. (I know at least one person who was tickled by this change ;)
- Updated the spec to describe marks in terms of "a sequence per executing function" instead of just program order. This is necessary to correctly describe function calls. Note @nhaehnle and @krzysz00 
- Renamed an existing subtarget feature `FeatureMemToLDSLoad` to match the internal names which say "vmem", and used this to guard the new mark/wait intrinsics. Note @arsenm 

https://github.com/llvm/llvm-project/pull/173259