Rewrite audio feature processors to accept torch tensors

Feature request

We would like to be able to torch.export the audio feature processors such as:

• GraniteSpeechFeatureExtractor
• WhisperFeatureExtractor used for Voxtral and Whisper

@ArthurZucker @zucchini-nlp @Cyrilvallez @larryliu0820

Motivation

This allows us to torch.export the whole E2E needed to generate output for Transformers audio models. Then in ExecuTorch, we can run everything without needing to write any custom C++ code for audio preprocessing.

Your contribution

Can help with code / provide context (?)