[llvm] [AMDGPU] Split large raw buffer offsets to SOFFSET where allowed (PR #207821)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> @arsenm for your thoughts (and whether you'd be opposed if this turned into a general addressing mode match)

Not sure what the question is here? We probably want something like this in selection, but given these are sourced from intrinsics, we probably should be doing heavier matching in the IR first 

https://github.com/llvm/llvm-project/pull/207821