Adding `safetensors` variant of this model

Replace with paper-clean model (punct-fixed dataset, focal loss)

Upload GUS-Net BERT model with optimized thresholds

initial commit