[New Model] HiFiGAN

[erogol]

This PR introduces the initial implementation of HiFIGAN in 🐸 territory.

• It is mainly adopted from the official implementation https://github.com/jik876/hifi-gan
• It also enables us to import officially released models with slight on-the-fly layer restructuring.
• audio.py is also updated to allow different log scales as it converts amplitude specs to DB. Traditionally we use log10 however some of the other projects (including HifiGAN) use log scaling. This update allows us to train compatible models with the other projects.
• Different than the original abstractive design of frog: TTS, we adopt HiFiGAN models with self-sufficient generator and discriminator model files to experiment with the pros and cons of this design choice.
• Optimizers are configurable from config.json file.
• Scheduler update steps are moved to the end of the epoch loop.

Thx 👑 @rishikksh20 who initiated the work and helped a lot to experiment and find the bugs.