[llvm] [ABI][AMDGPU] Add AMDGPU target ABI classifier to the LLVM ABI library (PR #220177)

[via llvm-commits]

skc7 wrote:

@yxsamliu 

> * large and kernel aggregates use `Indirect` instead of `IndirectAliased`;

Created pre-requisite PR: #225005 to handle this. 

> * non-trivial C++ returns use address space 0 instead of private address space 5;

Non-trivial C++ records return indirectly in the flat AS 0 via `maybeCommonClassifyReturnType`.
Large trivial aggregates use the private AS 5. Added the `NonTrivialRecordReturnIsIndirectFlat` test.

> * `struct { short a[2]; }` is coerced to `[2 x i16]` instead of `i32`.

Created pre-requisite PR: #224994 to fix the bug where a multi-element array member was incorrectly reduced to a single element.

https://github.com/llvm/llvm-project/pull/220177