Attention-guided Self-reflection for Zero-shot Hallucination Detection in Large Language Models