Improving Self-supervised Pre-training via a Fully-Explored Masked Language Model