On the Expressiveness of Approximate Inference in Bayesian Neural Networks