Generate: add model class validation

[gante]

What does this PR do?

Fixes #18210

This PR adds model class validation at the start of generate (all model classes inherit GenerationMixin, but few can use generate()). It also adds an exception that attempts to redirect the users to the right classes.

[sgugger]

Thanks for fixing!

[patrickvonplaten]

(nit) Wondering if it could make sense to make a really nifty error message here where we check if any class of the model type of the given class in in one of the auto classes below - maybe over-engineered though

[patrickvonplaten]

Sounds good! Wondering if it could make sense to make the error message a bit more detailed E.g. if someone uses BertModel and gets an error that one should use one of the following AutoModel classes, I'm not sure how helpful this is). Could be super cool that if one uses BertModel and gets as an answer that BertForCausalLM supports generate it could be a nicer UX. But maybe over engineered.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[gante]

@sgugger @patrickvonplaten I've requested a re-review of this PR. As per @patrickvonplaten's suggestion, the PR was upgraded to contain the exact class the user should use in the exception (as opposed to pointing to all generate-compatible auto classes).

In the process of building it, I've noticed that the previous version of this PR was incorrect anyways -- PT and TF had a default prepare_inputs_for_generation, so we couldn't rely on its existence. Only 1 model was using this default, so I removed it and implemented it in the missing model. The default prepare_inputs_for_generation was a public method, but since this PR blocks the use of generate() with classes that are not intended to be used with it anyways, removing the public method should have little impact. Nevertheless, it is a point to consider in the review!

---

Here's an example with the current version of the PR:

>>> from transformers import AutoModel
>>> model = AutoModel.from_pretrained("distilgpt2")
>>> model.generate("foo")
TypeError: The current model class (GPT2Model) is not compatible with `.generate()`, as it doesn't have a language model head. Please use one of the following classes instead: {'GPT2LMHeadModel'}

[LysandreJik]

Looks good to me! I haven't tested thoroughly but from a few quick tests it seems to work well.

Thanks @gante!

[sgugger]

Oh missed the notification that re-requested the review. Still LGTM!