arxiv RAMM: Retrieval-augmented Biomedical Visual Question Answering with Multi-modal Pre-training