Multimodal hugging Face wrapper, misc improvements, model debugging utils

[jlamypoirier]

✨ Description

• Add HuggingfaceMultiModalModelForCausalLM wrapping multimodal models for hugging face, following the llava format.
• Integrate the content of HuggingfaceBaseModelForCausalLM into HuggingfacePreTrainedModel and generalize to arbitrary inputs.
• Rework output_hidden_states into an extensive debugging utility using the existing DebugLayer. When calling the model, one may "request" the model to output specific hidden states by providing a list of names in kwargs["output_hidden_states"] (output_hidden_states in hf wrapper). The matching hidden states (using regex) will be returned in kwargs["hidden_states"]. This is still experimental but already helped a lot with degugging. Ex:

  >>> model_fast_llm(test_input, pixel_values=pixels,output_hidden_states=["vision_encoder.encoder.0.mixer", "head.logits"])
  CausalLMOutputWithPast(loss=None, logits=tensor(...), past_key_values=[], hidden_states=
  {'vision_encoder.encoder.0.mixer.query_rotary_input': tensor(...),  'vision_encoder.encoder.0.mixer.key_rotary_input': tensor(...), 
  'vision_encoder.encoder.0.mixer.query': tensor(...), 'vision_encoder.encoder.0.mixer.key': tensor(...), 
  'vision_encoder.encoder.0.mixer.value': tensor(...),  'vision_encoder.encoder.0.mixer.context': tensor(...), 
  'vision_encoder.encoder.0.mixer': tensor(...), 'head.logits': tensor(...)}, attentions=None)
  

• Replace the patch "convolution" by a simpler linear layer.
• Add support for linear layers without input gradients (ex. vision embeddings)
• Fix patch ordering in get_patches_from_images
• Add missing causal and cross_document_attention in llava conversion.