Core implementation of Jina XLM-RoBERTa

This implementation is adapted from [XLM-Roberta](https://huggingface.co./docs/transformers/en/model_doc/xlm-roberta). In contrast to the original implementation, this model uses Rotary positional encodings and supports flash-attention 2.

### Models that use this implementation

to be added soon


### Converting weights

Weights from an [original XLMRoberta model](https://huggingface.co./FacebookAI/xlm-roberta-large) can be converted using the `convert_roberta_weights_to_flash.py` script in the model repository.