Weak-to-Strong Backdoor Attack for Large Language Models