Implementation of ROBERT on top of Transformers and Flux

- This implementation is the same as `Transformers.Bert` with a tiny embeddings tweaks.
- [RoBERTa](https://arxiv.org/abs/1907.11692) has the same architecture as BERT, but uses a byte-level BPE(implemented in `BPE.jl`) as a tokenizer (same as GPT-2) and uses a different pre-training scheme.
- RoBERTa doesn’t have token_type_ids, you don’t need to indicate which token belongs to which segment. Just separate your segments with the separation (or ``)

we can also wrapper [Camembert](https://camembert-model.fr/) (or the french version of BERT) around RoBERT.