[llvm] [AMDGPU] Change SGPR layout to striped caller/callee saved (PR #127353)

[Christudasan Devadasan via llvm-commits]

cdevadas wrote:

This patch would improve the codegen with less number of SGPR spills for heavy workloads involving device calls. May be run the perf PSDB as well? That would give us some initial numbers. You can CP this PR to the staging compiler and then launch the perf PSDB.

https://github.com/llvm/llvm-project/pull/127353