Discrete (mixed) density estimators

[michaeldeistler]

For discrete (mixed) parameters.

Can probably recycle a lot of code from MNLE.

[janfb]

Yes, MNLE implements a mixed estimator with a categorical distribution.
It would be nice to extend this to discrete flows as well.

[gmoss13]

Related to #968

[coschroeder]

happy to work on this during the hackathon

[janfb]

The MNLE classes are now refactored to match the API of the other build functions, including z-scoring and embedding nets. Thus, in principle, one can now also use the MNLE setup for posterior inference. However, it allows only for a single discrete column in x (theta).

An autoregressive density estimator on top would be the solution, see #1112