On Teacher Hacking in Language Model Distillation