Change model loading, add HiFigan model.

Adding a downloadable HiFiGan artefact.
HiFiGan credit: https://github.com/jik876/hifi-gan

Forward Model

Added Forward Model based on FastSpeech, with scripted training on LJSpeech.
Fixes to previous model and improvements to layers structure.
Samples, colab notebooks and pre-trained models are available.

Autoregressive Transformer

Implementation of a Transformer phonemes to spectrograms model.
Scripted training on LJSpeech.
Working pre-trained model.