CO-BED: Information-Theoretic Contextual Optimization via Bayesian Experimental Design