Code cleanup

[fdreyer]

• Add the n_avg step and its inversion to the preprocessing pipeline, with some new method to get the image between the n_avg and the rescaling step
• Remove unused options like deterministic (perhaps also mnist stuff?), update the default runcards with results of scans
• Check that the saving/loading mechanism works for all the important models (LSGAN, WGANGP, VAE,...)
• Create a generation script that loads in a model + preprocessor and generates a given number of images, then either saves a series of plots or a numpy array
• Remove generated image sample from the default output, add more options for output plots
• Update the README with more detailed instructions for installation/download