Add statistics for N_eff over inferred weights, including worst-performing layers, in addition to r_hat