The Complexity of Bayesian Network Learning: Revisiting the Superstructure