Test that inferred posteriors of latent values concentrate at the true values

[emilyfertig]

• Take a single sample from an HIRM with a noisy relation and an unobserved base relation (n records of observed data)
• Incorporate that sample into a new HIRM
• Call transition_latent_value multiple times to generate posterior samples
• Check that the posterior concentrates at the true (sampled) value, i.e. the mean is close to the true value and the variance is low (or for Categorical e.g., the posterior is close in L1 distance to a deterministic distribution with all of its mass on the true value)