Attention Tracker: Detecting Prompt Injection Attacks in LLMs