Enable unsloth as training backend

[FilippoBoni1921]

Description

This PR has the goal to select unsloth as training backend in order to improve GPU usage efficiency

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[willccbb]

Definitely want to have a nice unsloth integration, and potentially open to adding something like this, but could we restructure this PR in a way where it's less breaking/intrusive?

We shouldn't add unsloth as required dependency, or remove the existing 'rl' dependencies (maybe use a new dependency group?); config classes should live in the same 'rl' subfolder and follow the established style/dev patterns.

In general, we're not planning on doing many feature changes to the included RLTrainer, which is intended as a "minimal demo trainer" that people can easily read + modify -- for performance + customizability, we generally recommend https://github.com/PrimeIntellect-ai/prime-rl, but it doesn't have all of the fancy Unsloth features for quantized models (and likely won't any time soon, below fp8 at least).

Also confused by some of the other changes, which seem unrelated (e.g. https protocol option)?

[FilippoBoni1921]

Thank you @willccbb for your reply! I know the PR was not very well structured. I did it fastly and started it as a draft.
I moved the UnslothConfig under the 'rl' subfolder and fixed the dependencies.
In my opinion creating a 'config' subfolder can be potentially useful for future extension eg a config dedicated to vllm args. For this reason I kept the UnslothConfig sperated in a different file, but let me know what you think.
In addition let me knoe what do you think about the methods to define the unsloth model. I tried to make them similar to the ones already existing. Unfortunately I had to insert them in a 'if/else' loop for the differences between 'trl' and 'unsloth'.
In conclusion https protocol option was a little experiment I have removed, that was unrelated.

Waiting for your feedback. Let me know if I can help in other issue for 'verifiers' or 'prime-rl'