Reducing Hallucinations of Medical Multimodal Large Language Models with Visual Retrieval-Augmented Generation