support / examples for roBERTa

Is there a way to run inference using a roBERTa model without converting it to onnx?

is it just a matter of implementing BERT, and modifying the tokenizer and using the embeddings for predictions like in the similarities example?

```rust
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(...)
tokenizer.decoder = decoders.ByteLevel()
tokenizer.post_processor = processors.RobertaProcessing(...)
```

The issue names no files or tests. Start by locating the existing BERT implementation and the similarities example, then review how tokenizer configuration is handled; done means documenting or demonstrating a working roBERTa inference path without ONNX conversion.

Written by the indexing model from the issue text.