Check model inputs - hidden states

[zucchini-nlp]

What does this PR do?

Fixes #40964

In most vision models the output.hidden_states are the hiddens right after encoder blocks, i.e. before the last layernorm. Therefore for these models output.hidden_states != output.last_hidden_state

Currently check_model_inputs assumes that last hidden state is the correct one to return which is true for language models only. This PR adds a kwarg for check_model_inputs which decides whether to replace last hidden state or not

TBH, i think the way it is done in LMs is the ultimate correct version and we probably need to "break" vision models. But I can't think of a way to obtain pre-norm last hidden states which are needed for some VLMs

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: aimv2, apertus, arcee, aria, audio_spectrogram_transformer, aya_vision, bitnet, blip, blip_2, cohere, cohere2, cohere2_vision, csm, deepseek_v2, deepseek_v3, deit

[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.