Training High-Level Schedulers with Execution-Feedback Reinforcement Learning for Long-Horizon GUI Automation