Cancer-Myth: Evaluating Large Language Models on Patient Questions with False Presuppositions