Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning