<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
Currently, the implementation for MLA V1 pads q, k, v to `head_dim` 256
to conform to early MLA kernel. But the new MLA kernel supports
`head_dim` that can't be devided by 128. Therefore we can remove those
unnecessary paddings to boost the performance
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: angazenn <zengyanjia@huawei.com>
Co-authored-by: angazenn <zengyanjia@huawei.com>
a970b27e2d