[WIP] Mini-batch training in VAE

[wehlutyk]

Work-in-progress pull request for #3, updated as work advances.

Feel free to comment or push other commits!

[wehlutyk]

Things left to do:

• [x] update layers.Bilinear to not use batch_size
• [x] update ae.build_p, and simplify the gcn-ae notebook accordingly
• [x] implement batches() to infer what it needs from the model
• [x] adapt the normalization of A for non-symmetric batch adjacency matrices
• [x] implement ae.Model.fit_feeding_generator()
• [x] scale each loss properly
• [x] implement dag.restrict() and update the notebook to show predictions