feature: Chat completion functions

[mudler]

Have a feature that mimics OpenAI functions: https://openai.com/blog/function-calling-and-other-api-updates.

From the documentation, it seems OpenAI trained the model to respond to the system prompt.

For LocalAI, I think there should be a minimum viable implementation that should work across models. While custom models could be trained to reflect a specific prompt that is injected, that would limit this functionalities to restricted models.

[mudler]

likely depending on #354

[mudler]

I've posted my updates on the upstream PR: ggml-org/llama.cpp#1773 (comment) to give early feedback on the feature

This depends on: ggml-org/llama.cpp#1773