[mlir][vector] Teach `TransferOptimization` to forward masked stores (#87794)

This only handles one case (that's fairly common in practice*), storing
a masked constant splat, then reloading again with the same mask and a
padding value that matches the splat.

* For SVE/SME (without peeling) this occurs when you have a
`linalg.fill` preceding a `linalg.matmul`.