Add support for modifying Processor Kwargs in modular

[yonigozlan]

What does this PR do?

Add support for modifying ProcessingKwargs, ImagesKwargs and TextKwargs in modular. It is currently needed for ColPali, and I think it's nice to have in general, as it could be useful for other multimodal models in the future. Added modular_new_kwargs_model.py to illustrate the need.

Also:

• Changed an error to a warning when an overwritten function isn't explicitly used in modular. This was making examples that used to work crash, and it seems a bit unnecessarily harsh 😅.
• Modified a bit the handling of relative imports to avoid false positives (not add "models" automatically if not present).
• Fixes a problem introduced in a previous PR, not sure which one, where a call to super() would be kept after copying the parent function, when it should be removed.

On that last point, I also added the modular examples to the default call of check_modular_conversion, as otherwise, the modular examples aren't automatically updated when changes are made to modular_model_converter, which makes it hard to trace when issues are introduced. It does make the recurrent checks such as fix-copies heavier though, so not sure if there's a better way to do this.

Who can review?

@ArthurZucker

[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.

[ArthurZucker]

Hey hey! Sorry, as you saw, @Cyrilvallez is refactoring the modular, this should now be fixed!