Designing Bayesian-Optimal Experiments with Stochastic Gradients