Basic pixtral support, paving the way for vision models 🖼️

[francoishernandez]

Conversion

Conversion is made to work with the mistral-community models, and was for now only tested with mistral-community/pixtral-12b.
The official mistralai models lack quite some information to be usable in our context. (HF checkpoints lack information as well, but it's more manageable.)

What works

The provided test_inference.py script in the pixtral recipe allows to run inference on a few examples (grabbed from the Pixtral blog post).
The configuration is using bitsandbytes quantization by default, to allow running on a 24G VRAM GPU (tested on 3090).
Some differences in a few methods (notably rope) lead to slight numerical differences with the HF implementation (as most of our models anyways).

What does not work (yet)

• finetuning was not tested
• batch inference is not supported (we could probably follow some of the discussion around here to make the necessary adaptations -- [Usage] Batch inference with Llava 1.5 haotian-liu/LLaVA#709)

What might need to be improved (future work)

• model definition (split out eole.models.model? allow for different adapter classes?)
• configuration / documentation
• proper configuration of specific image tokens
• HF conversion script (not only related to this work)
• dataloading is a bit hacky and might benefit some refactoring/simplification
• support more Llava-like models?

[francoishernandez]

Recent changes (mostly #171) broke a few things here, I grabbed the changes and created a clean commit from there.

Also, "model-level rope" might not the best here because we have different kinds of rope for the vision encoder and the text decoder. So for now, the vision encoder rope is defined within the encoder code, while the "model-level rope" is the text decoder one.

[francoishernandez]

The initial goal was to extend this PR with more llava-like vision architectures. This could have facilitated finetuning experiments with smaller models for instance.
But the various llava models are quite a mess with lots of vision encoder/adapter variants notably.
(For now, the finetuning codepath has only been tested "naïvely" by reducing the number of layers to fit in VRAM and making sure the code runs.)
So let's merge this, which allows to play with pixtral-12b inference, and address these topics (+ server support) in subsequent PRs.