Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench