Neutralizing Backdoors through Information Conflicts for Large Language Models