SceneTAP: Scene-Coherent Typographic Adversarial Planner against Vision-Language Models in Real-World Environments