From Pixels to Predicates: Learning Symbolic World Models via Pretrained Vision-Language Models