ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model