Beta binomial deconvolution

[pawel-czyz]

This PR adds the utilities for doing joint deconvolution and fitness estimation with beta-binomial model based on the count data.

It includes the utilities to fit such models and the examples on simulated data, using HMC or variational inference to infer the selection dynamics parameters.

Current limitations (to potentially be addressed in different PRs):

• It seems that we have some amount of code duplication right now. The changes are mostly self-contained (in a subpackage), so maybe it's not a big issue.
• It's currently only on simulated data.
• Some VI examples there (e.g., the mean field or the IAF) I tried gave rather wrong results, compared with HMC, if too simple family is used or it's not run for long enough.