CoCA: Regaining Safety-awareness of Multimodal Large Language Models with Constitutional Calibration