[llvm] [AMDGPU] Add heuristic to set CarriedLatencies (PR #224057)

[Krzysztof Drewniak via llvm-commits]

https://github.com/krzysz00 commented:

Not commenting on the merits, but what prevents us from having this be the default behavior? Compile time impacts?

https://github.com/llvm/llvm-project/pull/224057