Sample then Identify: A General Framework for Risk Control and Assessment in Multimodal Large Language Models