LLM-Powered Explanations: Unraveling Recommendations Through Subgraph Reasoning