More options for continual learning

[janfb]

This goes back to #51. #51 was closed because it relates only to where to reset to when retraining from scratch.

This issue is about providing more choices for the user how to continue training of the posterior across rounds. Pasting comments from #51:

@meteore wrote:

• improve docstring
• provide entry point for a more general choice of what weights to update on for a given round.
• explore interactions with pytorch-lightning

@janfb wrote:

In the future there will be different ways to transfer learning between rounds,

• continue training
• bayesian NN

we should add a function that decides between those, e.g., prepare for pytroch lightning.