[llvm] Reapply "StructurizeCFG: Optimize phi insertion during ssa reconstruction (#101301)" (PR #114347)

[via llvm-commits]

ruiling wrote:

@yxsamliu has helped with the debugging work of the original regression. The root cause is not the change itself, but hit certain deep issue in the register allocation of AMDGPU backend. With the recent fix and improvement in the register allocation part, we should be safe to land this again.

https://github.com/llvm/llvm-project/pull/114347