Exploring the Transferability of Visual Prompting for Multimodal Large Language Models