Integrate chat templates

[amakropoulos]

HuggingFace has enabled chat templates alongside the model definitions.
"They specify how to convert conversations, represented as lists of messages, into a single tokenizable string in the format that the model expects."
This is part of a WIP PR from llama.cpp.
Once this is merged and integrated in llamafile it will be also incorporated here.

[amakropoulos]

Chat templates seem to be blocked due to lack of simple c/c++ implementation.
jinja2 is a python engine and probably the only jinja compiler in c/c++ is Jinja2Cpp which is quite heavy.

references:
https://www.reddit.com/r/LocalLLaMA/comments/1afweyw/quick_headsup_about_using_codellama_70b_and/
ggerganov/llama.cpp#4216 (comment)