[CODE IMPROVEMENT] Deprecate RLHF with PPO in favor of DPO

[pascal-pfeiffer]

🔧 Proposed code refactoring

Deprecate RLHF with PPO in favor of DPO.

Motivation

RLHF with PPO involves an extra step to train a reward model (see #175).
For good results, the reward model needs to be of very high quality and tuned on a large amount of data. While OpenAI and Meta have successfully used PPO in their pipeline, the OS community struggled to get good results. Probably due to the lack of good training data for the reward model.
DPO (#530) is a technique that allows to train with Human Feedback in a more stable way on good+bad sample pairs and without the need for an additional reward model. It is also much quicker to train as the generate steps are skipped.