Hi Andrey,
If I can share my experience with the second part of the questions. I did also ask myself if there is a difference between the existing software and PPLs, and in Gelman’s book, it is mentioned that there are three ways of assigning probabilities: making subjective assessments, using empirical probabilities based on observed data, and constructing a parametric probability model.
People use PyMC3 mostly for writing a parametric type of model. However, it is possible to do all the other ways similar to the BayesFusion. You can check the forum for more examples if you are interested to model your data in that way. For example, this post Bayes Nets, Belief Networks, and PyMC
Good luck