World-To-Image: Grounding Text-to-Image Generation with Agent-Driven World Knowledge