Span-Level Hallucination Detection for LLM-Generated Answers