Sparse Progressive Distillation: Resolving Overfitting under Pretrain-and-Finetune Paradigm