XLMRoberta support

[Oliver-Y]

Added support for XLMRoberta model. Tested on Multilingual E5 embeddings model. It seems in the tokenizer.json of E5 a preprocessor is used but since llama.cpp doesn't support SPM preprocessors yet I put a simple workaround right before the SPM tokenizer call.

This is my first time contributing so would love feedback of any form!

• [x] I have read the contributing guidelines
• Self-reported review complexity: Low-Medium
  • [x] Low
  • [X] Medium
  • [ ] High

[Oliver-Y]

Might be a bug w/ tokenization. Going to take a look first

[Oliver-Y]

Redundant to #8658 so closing