[llvm] [AMDGPU] Lower `llvm.amdgcn.queue.ptr` instrinsic to using implicit kernel argument if feasible (PR #103490)

[Austin Kerbow via llvm-commits]

https://github.com/kerbowa commented:

It looks like the intention of @changpeng COb5 patches was that we should never allocate or access the user SGPR queue_ptr with COv5 https://github.com/llvm/llvm-project/blob/f3ca3ef4fca1c3f1b81278bc3823a791fb5efe2b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp#L2413.

Otherwise if we go with Matt's suggestion to make the intrinsic mean the preloaded SGPR then the allocation needs to be fixed.

https://github.com/llvm/llvm-project/pull/103490