Add support of audio transformers

[g-prz]

This PR proposes an integration of audio transformers into outlines framework (#1270)

[g-prz]

Hey @rlouf 🙋‍♂️
I'm not sure, but it seems that the runner disk is full and qwen2-audio files are taking a bit too much space, I can try to find a gguf version or maybe a smaller model? Wdyt?

[rlouf]

We generally use HF's internal testing models which are much smaller, see the other tests. Can you try to find if there is one for audio?

[g-prz]

Hey @rlouf
Running into something funny, I have test_generate_text[model_transformers_audio-beam_search] that fails, after investigation the instantiation of the beam_search sampler with default param beams to 1 leads to a greedy sampler in hugging face transformers which leads to not calling _beam_search down to _reorder_cache that is not implemented for Qwen2...

To fix it, we could split the parametrization of test_generate_text with beam_search separated from greedy and multinomial so that we can pass more than one head to beam search and not fall into greedy sampler case

[g-prz]

Hey @rlouf
I updated test_generate_text
I have 95% coverage, the missing parts are in case of import errors (not sure we want to modify the test env to have the coverage) or processor with no tokenizer, to what extend do you want to increase the coverage?
Otherwise good for review on my side