Learning Bayesian Metanetworks from Data with Multilevel Uncertainty