convert MonetAutoencoder to a variational autoencoder

[jeff-regier]

• add a term to the loss function that penalizes the KL divergence between the posterior and the prior
• show that the SpatialVAE has a higher ELBO value than a TrivialVAE (based on TrivialAutoencoder)