Add the possibility of a backend for audio decoding

### Feature request

Add the possibility of a backend for audio decoding. Before version 4.0.0, soundfile was used, and now torchcodec is used, but the problem is that torchcodec requires ffmpeg, which is problematic to install on the same colab. Therefore, I suggest adding a decoder selection when loading the dataset.

### Motivation

I use a service for training models in which ffmpeg cannot be installed.

### Your contribution

I use a service for training models in which ffmpeg cannot be installed.