Fix {image/video/audio}_token_ids in ProcessorMixin

[hmellor]

#45493 converted these attributes to properties. This breaks custom models which set {image/video/audio}_token_ids directly because:

• There is no setter, so these properties are read-only
• The getter will only ever read from image_token_id

This PR maintains BC by:

• Adding a setter for these properties
• Updating the getter to to first check if {image/video/audio}_token_ids has been explicitly set

[github-actions]

CI Dashboard: View test results in Grafana

[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.

[zucchini-nlp]

Thanks, just to be safe let's check that these private properties aren't getting serialized in a dict when we save

[hmellor]

>>> from transformers import ProcessorMixin
>>> processor = ProcessorMixin()
>>> processor.image_token_ids = [0, 1, 2]
>>> processor.image_token_ids
[0, 1, 2]
>>> processor.to_dict()
{'processor_class': 'ProcessorMixin'}