Improving Commonsense in Vision-Language Models via Knowledge Graph Riddles