Make docstring examples more stable

[jpreszler]

Several of the docstring examples have numeric output that changes too much leading to doctest failures.

We should change the data and models in those examples (particularly PrePostFit and RegressionDiscontinuity) to have more stability even if the examples become unrealistic or 'silly' - they will illustrate basic usage and ensure documentation is up to date with code, leaving the main docs for real examples and detailed instruction.

The main goal is that doctests should always pass unless there's been a clear change in how a function works.

[jpreszler]

I'll try to get this done over the next week.

[drbenvincent]

I was too hasty in merging #285.

Looking into it in more detail, it looks like we are getting non reproducible results. This seems to be because we are only ingesting the sample kwargs into pm.sample. So the random_seed kwarg is not being passed into pm.sample_prior_predictive or pm.sample_posterior_predictive.

We can't simply unpack the provided kwargs into these function calls because they don't accept all the same kwargs as pm.sample.

I'll submit a PR for this.