[llvm] [AMDGPU] Prefer lower total register usage in regions with spilling (PR #71882)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

Directly calculate excess pressure in `GCNPressure.less()` to remove patch dependence and decouple. 

We should still separate the `RegionsWithExcessRP` flag in scheduler as it reduces compile time and improves scheduling in regions with spilling, but decoupling is preferred.

https://github.com/llvm/llvm-project/pull/71882