Musketeer: Joint Training for Multi-task Vision Language Model with Task Explanation Prompts