From Novice to Expert: LLM Agent Policy Optimization via Step-wise Reinforcement Learning