Natural Language as Policies: Reasoning for Coordinate-Level Embodied Control with LLMs