Implement options for selecting layers to apply LoRA to during training

[viktorhargitai]

Implements a feature for specifying the layers of the model that LoRA should be applied to during training. This enables further training memory use reductions, and additional QLoRA training experiments (including but not limited to the reproduction of relevant experiments from the QLoRA paper).

The layer selection is done with the lora_modules str argument. Its value can be:

• a RegEx pattern for exact matching arbitrary layer names,
• the default value or 'all' selects all linear transformer block layers (i.e. no difference from the last commit)
• 'attention', which selects the attention layers only,
• 'ffn', which selects the feed-forward ones.
  (The latter 2 aim to reproduce ablation experiments from the paper, but work for e.g. Falcon models too, not just LLaMA.)

While the "lora_modules" argument was already in the guanaco training scripts, it was ignored because only the selection of all linear layers was implemented in this repo.