PREFER: Prompt Ensemble Learning via Feedback-Reflect-Refine