Co-authored-by: Gabriele Sarti <[email protected]>

Introduced in this PR:

- Decoder-only model support: Models like GPT-2 and Galactica can now be used through Huggingface without any code change (just specify model identifier in load_model).

- Model architecture classes: Separating architecture and framework responsibility in AttributionModel introducing framework-agnostic architecture classes (EncoderDecoderAttributionModel, DecoderOnlyAttributionModel) to handle input processing and attribution params formatting. Final classes will be children sharing one framework and one architecture (e.g. HuggingfaceEncoderDecoderModel).

- Fix partial target-side attribution: These were added at the beginning but didn't keep up with newest features, resulting in wrong slices when non-default attr_pos_start and attr_pos_end parameters were passed. These should now be fully fixed and functional when attribute_target=True.

- Enforcing a batch size of 1 when dealing with batched constrained generations with decoder-only models, or batched generations with custom start-end positions to avoid padding issues.

- Preliminary MPS GPU support in Torch (to be further tested)