Probabilistic Reach-Avoid for Bayesian Neural Networks