Small additions: working point script, customisable matplotlib backend, fix for evaluating non-freshly-trained models

This MR bundles three small changes:

1. adds a script for determination of the efficiency working points based on the tagger scores stored in the ntuples
2. introduces the option to set the matplotlib backend (now defaulting to Agg which is more fortunate for plotting in bash mode as opposed to tkinter)
3. adds a fix for evaluating a non-trained model (see https://umami-docs.web.cern.ch/WO_trained_model/) to use a dummy epoch of 0.