Integrate DeepSpeed's new HF interface for sequence parallelism

[ronald-d-rogers]

What does this PR do?

This PR adds changes to support DeepSpeed's updated mesh device interface for sequence parallel training, introduced here:
microsoft/DeepSpeed#5774

It adds flags to DeepSpeedPlugin and HFDeepSpeedConfig which can be used in tranformers to determine the state of sequence parallelism.

Note that in sequence parallel the inputs must be sharded such that only the each rank in the sequence parallel groups gets only its shard of the sequence.

If you look at the commit history, you will see that I toyed with doing this in accelerate itself, but ultimately determined that the sharding must be done in the trainers, right before the call the model(**inputs), otherwise libraries with custom trainers (i.e. `trl) would have to refactor the way their trainers prepare data.

Currently, this must be merged before the corresponding ds-ulysses PR in tranformers is merged.

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,
  Pull Request section?
• 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.

[SunMarc]

cc @XuehaiPan

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.