Modifications for large number of utterances and reducing data size

[Patrick LT]

I was trying with my own dataset for preprocessing and training the WaveRNN only.
Following modifications might be necessary for maintaining the training time and reducing the data size:

1. utils/dataset.py line 53
   num_workers is necessary to be adjusted as 1 per ~840 utterances, e.g., if the total data is 8400 or 9000, set to 10

2. preprocess.py line 47
   write quant as np.int16 type to reduce the size by a factor of 4, and to reduce I(/O) time to a lesser extent

3. dataset.py line 78
   label as np.int16; not as critical as the 1st/2nd points