Optimal experimental design via Bayesian optimization: active causal structure learning for Gaussian process networks