Detecting Backdoor Samples in Contrastive Language Image Pretraining