Refactor a lot: add several new features and misc improvements

[r9y9]

Things will be addressed are as follows:

• [x] Upsample network based on https://distill.pub/2016/deconv-checkerboard/
• [x] Wide context window for conditioning inspired by https://github.com/fatchord/WaveRNN
• [x] Remove tensorflow dependency
• [x] Use https://github.com/espnet/espnet style mel-spectrogram
• [x] train.py various fixes
• [x] dataset various fixes
• [x] Batch decoding
• [x] Kaldi-style egs (maybe)
• [x] Gaussian WaveNet #157
• [x] DC removal filter as same as https://github.com/kan-bayashi/PytorchWaveNetVocoder.
• [x] Fix tests/travis
• Include scripts to install
• [x] README update
• [x] cmu arctic is no longer (explicitly) supported, need to add doc for it

Later

• Pre-trained model for ljspeech
• New release after merging

Fix #157 Close #135