TCOD: Exploring Temporal Curriculum in On-Policy Distillation for Multi-turn Autonomous Agents