Fix enable memory efficient attention on ROCm

[tenpercent]

What does this PR do?

Fix enable memory efficient attention on ROCm while calling CK implementation

https://github.com/facebookresearch/xformers/blob/main/xformers/ops/fmha/ck.py#L157

Bug cause: the only supported dtypes by the op are float16 and bfloat16, and by default the input tensor is initialized as float32, which prevents using memory efficient attention on ROCm
Bug fix: pick a supported dtype when a custom attention op is passed as an argument
Note: attention op is not a Callable but a pair of (forward, backward) op implementations
https://github.com/facebookresearch/xformers/blob/main/xformers/ops/fmha/__init__.py#L202
https://github.com/facebookresearch/xformers/blob/main/xformers/ops/fmha/common.py#L469

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.