Toward Understanding Privileged Features Distillation in Learning-to-Rank