Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and Empirical Findings