ProKeR: A Kernel Perspective on Few-Shot Adaptation of Large Vision-Language Models