Calibrating Verbal Uncertainty as a Linear Feature to Reduce Hallucinations