Scalable Bayesian optimization with high-dimensional outputs using randomized prior networks