Upgrade data mixer, deps, and scripts

[lewtun]

This PR is a major refactor to bring the project in line with recent versions of TRL and to enable support for the SmolLM3 recipe. Some things may differ slightly from before (e.g. the train/test splits from the data mixture), but overall functionality should be the same.

TODO

• Test SFT and DPO work with Zephyr
• Fix the SmolLM3 recipe to use public models and test each step runs
• Refactor ORPO and test it works
• Update all remaining configs with the new dataset mixer
• Update unit tests

[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.

[lewtun]

cc @BramVanroy I have removed your CPT recipe for now as the mid-training dataset isn't compatible with this version of datasets (loading scripts deprecated) and the SFT model of your seems to have been deleted. Happy to add it back once the artifacts are back online!