Update TRL documentation

[qgallouedec]

Hey, TRL maintainer here 👋

This PR updates everything related to TRL to be more aligned with the current TRL API and the new best practices.

• QLoRA guide:
  • the dataset is now prompt-completion, so TRL masks the prompt itself and the 70-line collator (skip_prepare_dataset, remove_unused_columns=False) goes away.
  • SFTTrainer gets the model id plus quantization_config and loads and quantizes the model itself, so from_pretrained and prepare_model_for_kbit_training go away too.
  • It still gets the -it tokenizer, because google/gemma-4-E2B ships no chat template. It's the tokenizer rather than the processor because, with the processor, TRL wraps string content into lists, and Gemma 4's template then adds a trailing space to the system turn that inference never sees.
• Full fine-tune guide:
  • SFTTrainer and the pipelines take model ids.
  • I dropped dataset_kwargs={"add_special_tokens": ..., "append_concat_token": ...}, which TRL removed in ⚰️ Remove deprecated huggingface/trl#3704 and now ignores.

Both pages ran end to end on an H100 with a 3-step training run.

[bebechien]

LGTM!