414 long running time of sample posterior predictive and eventual death by oom

[AlexanderFengler]

• posterior predictive now has a safe_mode that chunks computations
• the n_samples argument was renamed to draws, and one can pass None | int | list | np.ndarray
• when running posterior predictive with kind='mean', the posterior naming cleans up rt,response_mean --> v
• prior predictives get assigned to .traces now, and naming is also cleaned up
• our sample_prior_predictive() will include the parent parameter as well now via an internal call to .predict()

[AlexanderFengler]

I will try to add a few more tests to this before merging.