LogicOCR: Do Your Large Multimodal Models Excel at Logical Reasoning on Text-Rich Images?