MMR: Evaluating Reading Ability of Large Multimodal Models (arXiv 2024)