Argparse integration and slight changes in train_model

Hey everyone,

I changed the structure in train_model a bit and added argparse to better work with the repo and minimize the number of entry points. To train a model you only need to hit python -m run.train_model --model AE --dataset test. Argparse currently handles the following options:

  --dataset {test,converging,w40,vw}
  --model {AE,ConvAE,VAE,VAECNN}