Visually Grounded Speech Models have a Mutual Exclusivity Bias