CoIN: A Benchmark of Continual Instruction Tuning for Multimodal Large Language Models