Use tf datasets instead of keras sequence

[zaccharieramzi]

This can be done in the same PR as #15 .

Basically the idea is to have a way cleaner data pipeline, and gain x3 time on training.

[zaccharieramzi]

This is dependent on tensorflow/io#745

[zaccharieramzi]

Actually, just re-tried with tf datasets using py_function and still using h5py, this results in a 1.5x acceleration compared to sequences for the pdnet training (see pdnet_af4_1568384763 vs pdnet_af8_1581355529). This also resolves #40 so will submit a PR solving both problems by replacing all sequences by datasets at least for training (will remove them from validation some time after).

[zaccharieramzi]

Partly adressed in #43 .
The validation should now use tf datasets as well to get rid completely of keras Sequences.