reupload BERT model without optimizer parameters in checkpoint

Created by: yurakuratov

Currently BERT models pretrained by DeepPavlov team contain optimizers params. It makes impossible to use these models with pytorch-pretrained-BERT without changing load_tf_weights_in_bert function.

current workaround to this issue: https://github.com/deepmipt/DeepPavlov/issues/863#issuecomment-497928718