[llvm] [AMDGPU] Flatten recursive register resource info propagation (PR #142766)

[Scott Linder via llvm-commits]

https://github.com/slinder1 commented:

Could we instead do a little constant propagation dataflow over the call graph? It should require less duplicated effort, and we should get the optimal propagation through the whole graph all in one go.

Or is it a goal to keep the "shape" of the expressions, so that e.g. `C.num_vgpr` in the newly added test case doesn't become the constant `71`?

https://github.com/llvm/llvm-project/pull/142766