[clang] [CIR] Wire AMDGPU into the call-convention lowering pass (PR #220197)

[Steffen Larsen via cfe-commits]

https://github.com/steffenlarsen commented:

The approach looks sound, though I wonder if we are doing ourselves a disservice by making target-specific divergence inside the functions. Might it make a hook similar to the target-info classes that will let us specialize the calling convention behavior based on the target?

Tag @adams381 for thoughts.

https://github.com/llvm/llvm-project/pull/220197