Incremental 'training'

[tsoporan]

Currently training happens on a frozen set of samples but this will change once the collector is running and pulling in new data. The current framework should support incremental training.

Basically this breaks down to keeping our word count histogram up to date and making sure we don't overwrite our existing data by using a reference such as an offset.

@lrvick