StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction — ThinkLLM