Test-time Backdoor Mitigation for Black-Box Large Language Models with Defensive Demonstrations