Multimodal Task Vectors Enable Many-Shot Multimodal In-Context Learning