Imitate then Transcend: Multi-Agent Optimal Execution with Dual-Window Denoise PPO