Use ML setup from HBW analysis to expand functionality

[jomatthi]

This PR expands the machine learning workflow of columnflow by introducing plots during training and more plotting task to study the performance of a trained network. It is taken from the hbw analysis and adapted to run in this analysis.

It is also more convenient to set up different ml models, there's even the basic infrastructure to perform a grid search with different parameters.

In addition, it was also necessary to fix the way events are categorised due to some earlier columnflow update introducing categorizer objects. This has been done already with the "production" categories in PR#9, but needed some further fixes. The new producer add_ml_cats has to also be called when producing histograms to make the ml categories available.

An exemplary workflow can be found in workflow_scripts/dnn_workflow.py.