feat: DeepSeek V3.2 Off-policy sequence masking

[casinca]

What does this PR do?

Fixes #4697

This PR aims to implement the Off-policy sequence masking from the DeepSeek V3.2 paper

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?
• Did you write any new necessary tests?

Changes

• added an helper _get_off_policy_mask staticmethod to compute the off-policy mask
• added logic in _compute_loss to inject the off-policy mask into the loss mask
• added logic in _compute_loss to inject the off-policy mask only to the surrogate loss
• added __post_init__ checks and some tests (including not allowing with Lieger loss atm)
• added docs

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.

[casinca]

• @qgallouedec 👋 I'd like to get your atttention on this commit: 861b490

  Initially I had made a mistake by injecting the off-policy mask into the loss mask early. So I had changed the loss mask to be computed upstream, once.
  Even if now, with the correct logic, I'm not touching the loss mask, I wanted to understand why is the loss mask recomputed multiple times? I don't see it being modified/overwritten anywhere downstream. In any case I can revert this specific commit, it won't change my implementation.

• Additionally if the PR is good, I could add a metric to log for users, something like a off policy sequence drop ratio, so that they aren't blind and also get a better feel to adjust the delta/off_policy_treshold hparam