Support return_tensors in audio chat templates

[zucchini-nlp]

What does this PR do?

Successor of #34275. Analogously supports vectorized output from audio LLMs. Currently we have only Qwen2Audio which needs to upload it template on the hub with minor changes instead of having the default_chat_template that is deprecated. Already opened a PR for that

From now on vectorized chat templates will work only for processors that are already uniform in terms of input kwargs, because we pass images, videos and audio and any of them can be None

This PR has done:

• Standardization on Qwen2Audio processor
• Some clean up on processor tests for audio models
• Audio chat template support
• Qwen2Audio default chat template has to be deleted when my PR on the hub is merged (TODO)
• Verify all multimodal instruct LLMs have standard processor kwargs otherwise they will fail to return tensors in chat template (TODO: VideoLlava is the only one that needs standardization now)

Qwen2Audio standardization raises one question:

• Currently if we have some defaults in tokenizer when init AutoTokenizer.from_pretrained(model_id, max_length=300, padding="max_length"), those defaults will also be used in audio feature extractor. So in this case the audio extractor will by default pad to max length of 300. I see that audio models usually pass same kwargs for tokenizer and feature extractor at call time, but I am not sure if we want it for the initialization

---

The docs will be updated as part of #35657 but the logic is same as with VLMs

messages = [
    {
        "role": "user",
        "content": [
            {"type": "audio", "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/glass-breaking-151256.mp3"}
            {"type": "text", "text": "What do you hear in this audio?"},
        ],
    },
]

inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True)
output = model.generate(**inputs, max_new_tokens=50)

[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.