FastSpeech2 + DaftExprt style Refernce Encoder with conditioning on Encoder, Variance Adapter and Decoder.

The issue during mismatch of mask and self attn shape is resolved, however, the results are not good. This is some unknown bug, despite the training logs showing promising results.

This version is the most stable FastSpeech2 version. This tag is the checkpoint to be used if we were to resume working on the vanilla fastspeech2 model.