PoTS: Proof-of-Training-Steps for Backdoor Detection in Large Language Models