Add LJU dataset, pretrained model and more

[ZDisket]

This PR includes:
1. LJSpeech Ultimate dataloader and pretrained Tacotron2 and vocoder (audio samples)
2. Forced Alignment Guided Attention Loss (FAL) from paper (WIP)
3. Multi-Band MelGAN-HF (mb melgan g + hifigan d), with both train from scratch and proven finetuning configurations. Also remove multiband_pwgan

The pretrained model was trained for 100k steps with regular training and then for 20k with technique described in 2. As far as I am aware, this would make TensorFlowTTS the first and only open-source TTS repo with a high sampling rate (44.1KHz) pretrained model available.

• Modify Tacotron2 dataloader to allow option for using premade attention masks
• Add alignment mask generation function from MFA durations
• Finish up READMEs

[dathudeptrai]

@ZDisket Is it finished ?

[ZDisket]

@dathudeptrai Almost done, except for two things: add instructions to Tacotron2 readme to use MFA with mfa_extraction for generating durations which will then be turned into masks, and a script to phonemize filelists as the LJSpeechUltimate processor takes in a file called filelist_p.txt with format for each line wavs/filename.wav|ARPA'd transcript. For example:

wavs/bztwyzpugz.wav|{ AH1 DH ER0 Z B IH0 L AO1 NG D T UW0 DH IY0 AE0 B EY1 Z AH1 V T UH1 R }
wavs/exlxkuhttq.wav|{ DH AH0 T D IH1 S AH0 P L AH0 N } , { IH0 N T EH1 N T AO1 N S AH1 T AH0 L T IY0 Z AH1 V L AA1 JH IH0 K AE1 N D M IY1 T AH0 } - { F IH1 Z IH0 K } , { W AO1 Z IH0 N D IH1 F ER0 AH0 N T T UW0 L IH1 T ER0 EH2 R IY0 F AO1 R M } , { AE1 N D S UW1 N B IH0 K EY1 M EH0 N K AH1 M B ER0 D W IH1 TH DH IY0 T EH1 K N IH0 K AH0 L }
wavs/aqzqsgksho.wav|{ T UW0 L IH1 B ER0 AH0 L AY2 Z M AH0 N AE1 S T IH0 K S T AH1 D IY0 Z }

I was thinking I could either do the PR without the filelist phonemizing tool and instead link it in the Tumblr blog post, or include it here, but I'm not sure.

[dathudeptrai]

@ZDisket can you use python -m black FILENAME for all your changed files ?

[ZDisket]

@dathudeptrai It's done. Now the processor will automatically phonemize strings when processing, and inference is as easy as this:

arpa_txt = processor.to_arpa(text)
ids = processor.text_to_sequence(arpa_txt)

I've tested everything I could think of so it should be ready to merge, although you are free to look at it if you've got the time. Four eyes better than two.

[dathudeptrai]

LGTM :D