AttentionRAG: Attention-Guided Context Pruning in Retrieval-Augmented Generation