* Move to tensorflow 2.0 and using keras for constructing neural network
* Adding specialize modules for Bayesian inference
* New distribution and distribution layers: `ZeroInflated`, `NegativeBinomialDisp`
* Add new statistical network: `MixtureDensityNetwork`, `DenseDistribution`, `DenseDeterministic` and more utilities for manipulating the output distributions in keras.
Tested against `tensorflow-2.0.0-rc1` and `tensorflow_probability-0.8.0-rc0`