AggTruth: Contextual Hallucination Detection using Aggregated Attention Scores in LLMs