Add support for Transformers audio models

[RobinPicard]

Addresses issue #1270

This PR turn the class TransformersVision into TransformersMultiModal that accommodates both vision and audio models. We should be able to use this same class for video models with minimal change, but I did not find a good model to test it out.

Something I'm unsure about concerns the keys of the dict provided as a prompt when calling a TransformersMultiModal model: I have the impression that they are standardized among transformers processors (text in all cases, images for vision models and audios for audio models), but I'm not sure of it. If it turns out they are not, we could remove the check on the keys of the dict and give responsibility to the user for providing what their processor needs.

I did not add model tests for audio models as the model I found (Qwen/Qwen2-Audio-7B-Instruct) is too heavy to run in the CI (and on some people's device).

[rlouf]

Suggested change

	"audios": audio_from_url(URL)
	"audio_files": audio_from_url(URL)

[rlouf]

Just a minor naming issue

[rlouf]

Suggested change

	"audios": Optional[Union[Any, List[Any]]],
	"audio_files": Optional[Union[Any, List[Any]]],

[RobinPicard]

I used audios because it's the keyword used by the Gwen2Audio processor (for the same reason I changed prompts into text). I think letting users choose whatever keyword their processor uses (apart from text that's shared by all) may be the best solution. We would still use inputs = self.processor(**model_input, padding=True, return_tensors="pt"). Another advantage is that it lets users provide optional kwargs for the processor call.

[rlouf]

If those are the name of the arguments in transformers so we can do inputs = self.processor(**model_input, padding=True, return_tensors="pt") let's keep them this way.