Driver Activity Classification Using Generalizable Representations from Vision-Language Models