Probing Language Models for Pre-training Data Detection