discrete token for audio resynthesis

[South-Twilight]

Here is the PR for audio resynthesis in discrete token:

1. We extend hubert_voc1 to token_voc1 and it can handle more models token;
2. We add f0 for training and inference when finding poor prounciation in singing;
3. We add multi-stream method including residual cluster and weight sum;
4. Using embedding feature of models is also allowed.

The following models have been validated in opencpop recipe: HuBERT, XLS-R, WavLM, MERT, Encodec.