allow for sound dataset to return multiple resample audio, since for training...

allow for sound dataset to return multiple resample audio, since for training the coarse transformer, the same audio needs to be resampled differently for the hubert or vq-wav2vec vs the soundstream coarse token ids