[llvm] [VPlan] Extract reverse mask from reverse accesses (PR #155579)

[Luke Lau via llvm-commits]

lukel97 wrote:

> We already include the cost of some VPInstructions after computing cost from the legacy cost model IIRC, we could do the same for reverse to get the more accurate estimates

Yes we can do that too, but would it better to split the cost model change off into its own PR? I think this PR can be NFC-ish otherwise.

https://github.com/llvm/llvm-project/pull/155579