[llvm] [VPlan] Perform optimizeMaskToEVL in terms of pattern matching (PR #155394)

[Luke Lau via llvm-commits]

lukel97 wrote:

I've unstacked this PR because I don't think the regressions are that serious, and can be fixed afterwards by #155383.

I think there's value in getting this in earlier as it should make both #146525 and https://github.com/llvm/llvm-project/pull/155579 much simpler to implement, cc) @Mel-Chen 

https://github.com/llvm/llvm-project/pull/155394