Utilize MonitoredTrainingSession for training

[gdj0nes]

Moving to this for training would have a number of benefits.

1. Improved support for model inspection via TensorBoard
2. Better recovery from failure and check pointing, automatic loading of most recent model from directory

[beckermr]

So 1 sounds interesting. For 2 sklearn's preferred serialization format is through joblib. How would this interact with that?