From Local Concepts to Universals: Evaluating the Multicultural Understanding of Vision-Language Models