MedFuzz: Exploring the Robustness of Large Language Models in Medical Question Answering