Text this: Detecting Hallucinations in Large Language Models via Internal Attention Divergence Signals