[CIR][AMDGPU] Track: AMDGPU call-convention lowering in ClangIR via the LLVM ABI library

**Summary:**
- Bring AMDGPU call-convention lowering to ClangIR by porting the classic CodeGen AMDGPU ABI classifier to the language-agnostic LLVM ABI library and wiring it into CIR's `CallConvLoweringPass`. This issue tracks the work.

**PRs:**
- [ ] #220579 - [CIR][ABI] Honor abi::ArgInfo::CanBeFlattened in the call-conv lowering
- [ ] #220177 - [ABI] Add AMDGPU target ABI classifier to the LLVM ABI library
- [ ] #220197 - [CIR] Wire AMDGPU into the call-convention lowering pass
- [ ] #220207 - [CIR] Support direct-in-registers aggregates and array coercions in call-conv lowering

This is still WIP and more PRs will be added as per requirement. I'll keep this issue updated as PRs land.

CC: @RiverDave @adams381 @steffenlarsen @ranapratap55