Probing Classifiers are Unreliable for Concept Removal and Detection