[llvm] [VPlan] Impl VPlan-based pattern match for ExtendedRed and MulAccRed (NFCI) (PR #113903)

[David Green via llvm-commits]

davemgreen wrote:

It makes sense to try and keep these cost models in-sync, but this might be doing better than the legacy one, I am still seeing a fair number of asserts. I've added two new tests to the same file, one has two reductions which have a shared sext node which in the old cost model would be cost-modelled but needn't have (multiple uses were difficult for it to deal with optimally). The other has a `ext(mul(ext, ext))` reduction that isn't handled by the old cost model properly. #115847 should hopefully help bring that case closer to this.

https://github.com/llvm/llvm-project/pull/113903