Reducing amount of training executables

Currently we have quite a bunch of training executables for dips, cond-dips, DL1, umami

It might be good to clean them up and have one executable like train.py which then takes as argument the tagger which wants to be trained and executes the dedicated training script