llama : add Refact support

[ggerganov]

This is a new 1.6B code model: https://huggingface.co/smallcloudai/Refact-1_6B-fim

We should look into adding support into llama.cpp similar as to how we did for Falcon and Baichuan:

• llm : add Falcon support #2717
• Feature: support baichuan serial models, by now, including Baichuan-7… #3009

I haven't looked into the architecture yet, but I'm hoping it is similar to GPT, which we already know how to handle OK thanks to the Falcon experience. I see it also uses Alibi which we should have support for, but we haven't tested extensively, so there might be issues there.

[Jipok]

Related: #3013

[discobot]

Let me attempt this :)