OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning