Fix sil (#22)

* Update infore dataset

* Use new textgrid data.

- Update download url and hash.
- Use sil instead of sp.
- Normalize audio to match hifigan preprocessing.
- Random dropout of tokens when training duration model to prevent overfitting.

* Load phoneme set from config instead from lexicon file.

This keeps the phoneme set unchanged even if the dataset or the lexicon file changed.

* use `jax.tree_map` instead of `jax.tree_multimap`.

* Better log file names

* Remove colab links in notebooks

* Fix `zero_silence_segments` script.

* Update pretrained models

v0.3.1: add fine-tune training for hifigan

v0.3.0: add word separator to avoid word mispronunciation, more prene…

…t dropout

v0.2: use HiFiGAN vocoder

update waveRNN pretrained model

minor fixes to make training work