FSDP-QLoRA doc updates for TRL integration

[blbadger]

Small documentation changes to maintain compatibility as the SFTTrainer class received updates as of TRL v0.13.0:

1. max_seq_length and dataset_text_field moved to become SFTConfig class arguments from the SFTTrainer, causing errors when calling the trainer if these arguments are retained.
2. The tokenizer argument is being deprecated in favor of the more general processing_class, which can be a tokenizer.

The doc currently does not define the training_arguments arg, but we can assume that this is a an instance of the SFTConfig class in which case max_seq_length and dataset_text_field would be defined there. I have removed these args from the SFTTrainer instantiation because of this, although these arguments could be defined in a config if that is preferable.

Swapping the tokenizer for processing_class deals with the second change. This modification is optional as the tokenizer kwarg is deprecated and swapped automatically, but will probably be useful going forward.

[Titus-von-Koeller]

Thanks @blbadger for you first PR and this proactive attitude towards fixing things to the benefit of the community ❤️ Really appreciated!

[blbadger]

Happy to contribute @Titus-von-Koeller !