* initial encodec * works * nits * use fast group norm * fix for rnn layer * fix mlx version * use custom LSTM kernel * audio encodec * fix example, support batched inference * nits
