Fix jetmoe model

[Cyrilvallez]

What does this PR do?

Fixes #31266
cc @ArthurZucker @amyeroberts

I am sorry I did not notice that JetMoePreTrainedModel got _supports_cache_class = True support while working on #30536.

To avoid the same bug in the future, any model getting _supports_cache_class = True in the future should change the line:

if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}

to

if inputs_embeds is not None and past_length == 0:
            model_inputs = {"inputs_embeds": inputs_embeds}

in prepare_inputs_for_generation() because checking past_key_values is None is no longer correct as empty but initialized DynamicCache can be passed.

Also to minimize code complexity, the if-else that checks if isinstance(past_key_values, Cache): (still in prepare_inputs_for_generation()) can be safely removed in favor of the case when this is True when _supports_cache_class = True as generate() will only use proper Cache classes in this case.
Some similar if-else could also be removed in upstream model architectures classes as well.

[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.

[amyeroberts]

Thanks for fixing, @Cyrilvallez!

Could you also remove the skips from the model tests which were added in #31266 as a patch?

I think the changes look OK, but let's have a second opinion from @ArthurZucker, as I'm not very familiar with the recent cache changes

[ArthurZucker]

we merged this for other models as well so LGTM.