Fix flaky test for multimodal LLMs

[Rocketknight1]

We have flaky test failures in tests/models/qwen3_omni_moe/test_modeling_qwen3_omni_moe.py::Qwen3OmniMoeThinkerForConditionalGenerationModelTest::test_generate_continue_from_past_key_values. The cause is that the logic in this test drops multimodal inputs (because it's passing past_key_values to generate instead), but it fails to drop audio inputs correctly. This causes flaky failures in LLMs like Qwen3OmniMoE, which I think occur when the audio encoder creates audio features that overwrite the embeddings for generated tokens (maybe when the tokens have the audio_token_id value?)

I'm not 100% sure of the cause, but the test fails 3% of the time without this fix and 0% with it, so I'm pretty sure dropping inputs correctly is the solution!

cc @zucchini-nlp

[zucchini-nlp]

Approving so you can merge after checking that the test passes for other audio models :)

[zucchini-nlp]

input_features is a quite common name across audio-models like whisper. Can you verify that the test passes for all models?
We can as well override it in qwen3-omni-moe test file, if it fails for other audio models

[Rocketknight1]

Yes! Dropping that input is intentional even when it's present on the model, unless it's the model_main_input. Because we're passing past_key_values to generation, we don't actually want to run multimodal inputs through the encoder again, and doing that is what caused the failure in the past.

[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]

Btw, the failing Bark was fixed on main :)

[Rocketknight1]

Ran the test for every model, and this change didn't cause any new failures, but did fix the Qwen3 test!