Improve query formatting & sizeing

[jepler]

huggingface transformers 4.34, which is quite new, has support for "chat templates" and can also tell you the size of a chat in tokens.

however, a lot of models don't have the required chat templates (yet?) and getting chat templates for some models (e.g., llama2) requires special permission even if a derived quantized model was not behind a signup wall.

Use this tech, or something like it, to replace the hard-coded query formatting of the llama_cpp backend and to improve the length of the query itself instead of having the hard-coded limit of 5 messages.

[jepler]

Note that llama_cpp backend should probably use the chat/completions API, in which case the template is applied on the server side and chap doesn't need to do anything: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template ggml-org/llama.cpp#11016

The local use of HF chat templates when calling to the huggingface inference API might still make sense.