Add an autoencoder

[sbdzdz]

Add a model that reconstructs the exemplar directly from a transformed image.

Changes:

• add the autoencoder class
• add a corresponding config
• remove the supervised VAE
• make the visualization callbacks more general