Test Rhat for the Gibbs kernels

[emilyfertig]

This consists of multiple tests that follow the same basic structure:

• Take a single sample from an HIRM
• Incorporate that sample into n new HIRMs
• Call the relevant transition method multiple times on each HIRM
• Use R-hat (Add a test utility that implements the R-hat statistic #131) to test stationarity/sampling efficiency

Test each of these components

• Transitioning entity cluster assignments, using the mean number of clusters as a metric
• Transitioning relation cluster assignments, using the mean number of clusters as a metric
• Transitioning latent values

The tests should be applied both to a case where the individual Gibbs kernel is transitioned repeatedly (with everything else held constant) and to a case where the whole model is transitioned.

[alex-lew]

This sounds ultimately good to have but maybe not highest-priority. (I have rarely used these diagnostics in the early stages of developing a model / inference algorithm, but it's possible there are good ways of doing so -- so I don't mean to discourage if we think it would be useful!)