Using google bert for RoBERTa

first, using convert_pytorch_checkpoint_to_tf.py convert wieght

second, edit config file

third, modified modelling:using_token_type=False start=2

Fourth, change bert tokenizer to roberta tokenizer