Refactor NeuralPosterior interface

[janfb]

There seem to be a couple of problems with the NeuralPosterior class as the main interface for all the posterior classes:

• it is treated as an abstract class, but it is not declared as such, e.g., what happens if one instantiates it and then calls sample?
• the doc string says it has a log_prob but it does not
• for a abstract base class, it has a lot of logic happening in this class, e.g., all the x-handling
• sample_batched and sample are quite related but have different signatures, e.g., the order differs.

How to fix this (?):

• get an overview how this class is used in the overall class structure
• define as a pure abstract class and move logic about x somewhere else, e.g., have a separate abstraction of x (Observation)?

@schroedk will probably be very helpful for working on this.