Fix: num_logits_to_keep in composite models

[zucchini-nlp]

What does this PR do?

Fixes issue from #33164 (comment).

We cannot assume that the model will have num_logits_to_keep in its forward, because an LM can be used as backbone in composite models. So the current PR sets num_logits_to_keep to None by default and adds it as a kwarg only when it is explicitly set to an integer by users

Plus, updated VLMs accordingly with num_logits_to_keep. Other VLMs like BLIP don't need this arg because they delegate the whole generation process to the LM.

cc @merveenoyan

[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.

[LysandreJik]

Understood, thanks @zucchini-nlp!

[LysandreJik]

Would it be possible to add a test to ensure no regression on the issue #33164 (comment) ?

[zucchini-nlp]

Oh, also cc @gante for generation :)

[zucchini-nlp]

@LysandreJik yes, I am working on enabling generation tests on all VLMs. That should be 2-3 PRs away, after we fix some issues after the latest refactoring on VLM processing code

We've overlooked many issues on VLM side due to the absence of fast generation tests and slow tests are usually not ran if modeling code is unchanged.