[X86] Fix EVEX compression for VPMOV*2M + KMOV with tied mask use (#198220)
When scanning uses of the mask produced by `VPMOV*2M`, we previously bailed out as soon as we encountered a write. For tied read/write mask instructions such as `KSHIFTR*`, which both read and write the same mask register, the pass could miss the use, fold the earlier `KMOV`, and erase the `VPMOV*2M` def even though the mask was still live.
Disclaimer: LLM came up with the MIR tests and explained this pass to me.
Fixes #198197