Fix torch only support for fast Processors

[Blaizzy]

What does this PR do?

Adds MLX tensor support to BatchFeature and removes the PyTorch-only restriction in fast image processing utilities.

Changes:

• Added is_mlx_array export to transformers/utils/__init__.py
• Implemented MLX tensor conversion logic in BatchFeature.convert_to_tensors() (feature_extraction_utils.py)
• Removed the hardcoded validation that restricted fast image processors to PyTorch tensors only (image_processing_utils_fast.py)

This enables using return_tensors="mlx" with vision models and fast image processors on Apple Silicon, which previously raised:

ValueError: Only returning PyTorch tensors is currently supported.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#create-a-pull-request),
  Pull Request section?
• Was this discussed/approved via a Github issue or the [forum](https://discuss.huggingface.co/)? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  [documentation guidelines](https://github.com/huggingface/transformers/tree/main/docs), and
  [here are tips on formatting docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
• Did you write any new necessary tests?

Who can review?

• @yonigozlan @molbap (vision models / image processing)
• @ArthurZucker (library / model loading)

[Rocketknight1]

cc @yonigozlan @molbap but not sure who to tag for MLX!

[Blaizzy]

@ArthurZucker has some context as well

[ArthurZucker]

LGTM, @yonigozlan is it fine for you as well?

[yonigozlan]

Hello @Blaizzy ! Sorry for the delay. I'm ok with merging this if it can improve compatibility with mlx! We don't have any mlx image processor in the library for now, but I'm guessing this is to use with custom image processors?
As discussed with Arthur, we'll move towards having different backends for image processors, so we could add one for mlx. However image processors are quite different across models, so we would need to have custom ones for each backend and models, as making image processing operation not library dependent seems difficult. Anyway these are issues for another PR ;)

[yonigozlan]

Not sure if these modifications are related to this PR or if they were added here by mistake?

[Blaizzy]

They were a mistake 😅. I was following the formatter feedback.

[Blaizzy]

We don't have any mlx image processor in the library for now, but I'm guessing this is to use with custom image processors

Thank you very much @yonigozlan! Yes, that will be the path I'll take in the meantime. Because for majority of models they can return mlx tensors from transformers but a few use torch and torch vision explicitly, that forces torch as dep.

[Blaizzy]

Suggested change

	if isinstance(value, np.ndarray):
	return mx.array(value)
	else:
	return mx.array(value)
	return mx.array(value)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: pop2piano

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=42824&sha=4adaad

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[yonigozlan]

Hey @Blaizzy ! I fixed some tiny formatting errors and moved the test to test_feature_extraction_utils.py‎. It would be great to get this merged soon as we're making progress on the PR to introduce image processing backends ;) #43514

However, I can't use mlx on my machine, and we don't have mlx support in our CI. Would you mind running the test (RUN_SLOW=1 python3 -m pytest tests/utils/test_feature_extraction_utils.py -k test_batch_feature_mlx_conversion) to make sure it passes?
Thanks!