Robust Knowledge Distillation Based on Feature Variance Against Backdoored Teacher Model