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

[via llvm-commits]

ruiling wrote:

> One very strong suggestion ... the "state" of this algorithm is all over the place. I think its worthwhile to consider making a single struct that contains all the information, with various methods that update its state. Kinda like a "locally designed" SSA updater.

I restructured the code to a separate function, looks a little bit better than before.

https://github.com/llvm/llvm-project/pull/101301