Stop training when a monitored quantity has stopped improving.

* store history information to json
* shuffle training data before fit
* include sample_weight
* separate load function to plot from previously trained setup
* weights and history are still stored after keyboard interrupt