Multi-Stage Knowledge Integration of Vision-Language Models for Continual Learning