II-Bench: An Image Implication Understanding Benchmark for Multimodal Large Language Models