Implement VSBC diagnostic from Yao et al.

[martinmodrak]

The Yes, but Did It Work?: Evaluating Variational Inference paper by Yao et al. proposes a less stringent diagnostic for variational inference than full SBC (as we know the approximation is almost always imperfect and SBC is thus eventually bound to fail). The main point is that we only test for bias, i.e. whether P(true_value < fitted_mean) = P(true_value > fitted_mean) across multiple simulations.

The implementation proposed in the paper seems specific to methods where we get exact mean, but the sample-based version should also work.

It might be neat to support this bias test, as it could be more generally useful in understanding if approximations are a problem or not.