[VLM] Update pixtral data collator to reflect latest transformers changes

[kylesayrs]

Purpose

• In transformers==4.48.0, the Pixtral processor was updated to not add an additional layer of wrapping for pixel_values ([PixtralLarge] Update Pixtral conversion script to support large format! huggingface/transformers#34801). This is more inline with how other processors handle multimodal inputs
  • Because previously the data_collator was being used to unwrap this unnecessary wrapping, attempting to quantize pixtral with transformers>=4.48.0 fails

Changes

• Update pixtral data collator to match latest transformers version
• Add comment for those who want to use transformers<4.48.0

Testing

• Ran pixtral example to completion, @shubhra ran pixtral large

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[brian-dellabetta]

👍