[Enhance] Optimize the performace of ms_deform_attn for MLU device

[Wickyzheng]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

The motivation of the PR is to improve the performance of ms_deform_attn on Cambricon MLU backend, especially for small channel cases.

Modification

MLU src code
Modify MLU src code of ms_deform_attn_mlu_kernel in directory mmcv/ops/csrc/common/mlu/ms_deform_attn_mlu_kernel.mlu.
PyTorch adaptation
Modify ms_deform_attn for PyTorch in mmcv/ops/csrc/pytorch/mlu/ms_deform_attn_mlu.cpp.

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repositories?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

Before PR:

• I have read and followed the workflow indicated in the CONTRIBUTING.md to create this PR.
• Pre-commit or linting tools indicated in CONTRIBUTING.md are used to fix the potential lint issues.
• Bug fixes are covered by unit tests, the case that causes the bug should be added in the unit tests.
• New functionalities are covered by complete unit tests. If not, please add more unit test to ensure the correctness.
• The documentation has been modified accordingly, including docstring or example tutorials.

After PR:

• If the modification has potential influence on downstream or other related projects, this PR should be tested with some of those projects, like MMDet or MMCls.
• CLA has been signed and all committers have signed the CLA in this PR.

[grimoire]

Test failed with log:

E                       torch.autograd.gradcheck.GradcheckError: Jacobian mismatch for output 0 with respect to input 0,
E                       numerical:tensor([[0.1410, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
E                               [0.0000, 0.1410, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
E                               [0.0000, 0.0000, 0.1410,  ..., 0.0000, 0.0000, 0.0000],
E                               ...,
E                               [0.0000, 0.0000, 0.0000,  ..., 0.3441, 0.0000, 0.0000],
E                               [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.3441, 0.0000],
E                               [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.3441]],
E                              device='mlu:0')
E                       analytical:tensor([[0., 0., 0.,  ..., 0., 0., 0.],
E                               [0., 0., 0.,  ..., 0., 0., 0.],
E                               [0., 0., 0.,  ..., 0., 0., 0.],
E                               ...,
E                               [0., 0., 0.,  ..., 0., 0., 0.],
E                               [0., 0., 0.,  ..., 0., 0., 0.],
E                               [0., 0., 0.,  ..., 0., 0., 0.]], device='mlu:0')

[openmmlab-bot]

Wickyzheng 您好！您在MMCV项目中给我们提的PR非常重要，感谢您付出私人时间帮助改进开源项目，相信很多开发者会从你的PR中受益。
我们非常期待与您继续合作，OpenMMLab专门成立了贡献者组织MMSIG，为贡献者们提供开源证书、荣誉体系和专享好礼，可通过添加微信：openmmlabwx 联系我们（请备注mmsig+GitHub id），由衷希望您能加入！
Dear Wickyzheng :
First of all, we want to express our gratitude for your significant PR in the MMCV project. Your contribution is highly appreciated, and we are grateful for your efforts in helping improve this open-source project during your personal time. We believe that many developers will benefit from your PR.
We are looking forward to continuing our collaboration with you. OpenMMLab has established a special contributors' organization called MMSIG, which provides contributors with open-source certificates, a recognition system, and exclusive rewards. You can contact us by adding our WeChat（if you have WeChat): openmmlab_wx, or join in our discord： https://discord.gg/raweFPmdzG. We sincerely hope you will join us!
Best regards！ @Wickyzheng