Mechanistic Exploration of Backdoored Large Language Model Attention Patterns