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

[Changpeng Fang via llvm-commits]

changpeng wrote:

> It looks like the intention of @changpeng's COv5 patches was that we should never allocate or access the user SGPR queue_ptr.
> 
> 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.

Right. We should only use the implicitarg field for queue_ptr for cov5. I do remember I have to update many LIT tests due to less sgpr usage. 

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