Fix the computation of KL divergence loss in Nash MD

[d-tiapkin]

What does this PR do?

The initial version of KL divergence in NashMD Trainer had a bug: the gradient of the given KL estimate did not correspond to the actual stochastic gradient of KL divergence between the current policy and the reference policy.

This PR intended to fix this issue by computing a REINFORCE-like estimate, taking into account that the current model generated the completion.

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? 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.
• Did you write any new necessary tests?

Who can review?

@kashif

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.