Refactoring `run` for additional flexibility about sbibm

[jan-matthis]

Thanks for the suggestion, Peter!

I agree that for additional analyses, it might be useful to be able to access the "trained object" (currently, depending on the algorithm, there is one trained object, or none, e.g., for some ABC methods). run does return not only predictive samples, but also the number of simulations that were actually performed (simulator.num_simulations; as a safety-check), and, if the algorithm supports it, the log probability of true parameters as an optional third return.

If you are up for it, I'd be glad to accept a PR that refactors the functions along the lines you propose and continue discussing there :)