Rethinking the BERT-like Pretraining for DNA Sequences