Mind vs. Mouth: On Measuring Re-judge Inconsistency of Social Bias in Large Language Models