Add support for Llama-3.2-vision & Resize image

[Blaizzy]

Adds support for Llama Vision, streamlined resize image and formated language model output.

• Test multi-image
• Add to trainer

Closes #60

[Blaizzy]

LGTM!

[jrp2014]

Some other models work with the example python code, but this model seems to need another parameter:

meta-llama/Llama-3.2-11B-Vision
Fetching 15 files: 100%|█████████████████████████████████████████████████████████| 15/15 [00:00<00:00, 36578.23it/s]
Fetching 15 files: 100%|█████████████████████████████████████████████████████████| 15/15 [00:00<00:00, 15076.58it/s]
/Users/xxx/Pictures/Processed/20241012-153359_L1010053.jpg
Traceback (most recent call last):
  File "/Users/xxx/Documents/AI/mlx/scripts/vlm/mytest.py", line 33, in <module>
    formatted_prompt = apply_chat_template(
                       ^^^^^^^^^^^^^^^^^^^^
  File "/opt/homebrew/Caskroom/miniconda/base/envs/mlx/lib/python3.12/site-packages/mlx_vlm/prompt_utils.py", line 140, in apply_chat_template
    return processor.apply_chat_template(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/homebrew/Caskroom/miniconda/base/envs/mlx/lib/python3.12/site-packages/transformers/processing_utils.py", line 1067, in apply_chat_template
    raise ValueError(
ValueError: No chat template is set for this processor. Please either set the `chat_template` attribute, or provide a chat template as an argument. See https://huggingface.co/docs/transformers/main/en/chat_templating for more information.

[jrp2014]

But the -Instruct version of the model works! It's very much slower than some of the other models on a 48Gb M3 Max.