Efficient Few-Shot Continual Learning in Vision-Language Models