[LARGE] Bring all recent updates from brrr – reducing dependencies

[thomwolf]

Goals:

• bring all recent updates from brrr
• organize configurations: we only use yaml as inputs
• clean up folder structure and reduce a bit the number of files, make naming more consistent
• remove dependency on datasets and transformers
• add simple scripts: run_train.py (former use_trainer.py), run_eval.py, run_generate.py
• remove dependency on lighteval
• remove dependency on s3
• remove logging on TB/WanB. Only keep stdout logging
• clean-up trainer (still changes in comments for now)
• add tiny_llama example script

[thomwolf]

Ok it now all run for the training. Example can be run by just doing bash ./examples/train_tiny_llama.sh

Maybe we can merge to use it as a base for the other PRs.

Cc @NouamaneTazi quite hard to review unfortunately sorry

[thomwolf]

Merging for now and we can iterate from main for further tweaks and fixes