Investigate sim_data in BetaGeoNBDRV

[ColtAllen]

The sim_data method is important for model evaluation because it generates synthetic data for prior/posterior predictive checks. Its current form was inspired by the equivalent function in lifetimes, which that library's developers admitted in a code comment is "hacky until I can find something better."

Reviewing the research, the bottom of p4 describes a Geometric distribution in the dropout process which may be a more appropriate approach. It would be interesting to see if this improves convergence results in the unit tests for BetaGeo.distribution_new_customer.