Support for jinja for custom chat templates

[Z1EMN1AK]

llama.cpp cli supports it but llama-cpp-python don't and we need it

[Legion-ale]

I go the same problem with jinja2 templates in ver 0.3.16.

Idk if it works even for your case, but I wrote a stupid fake_formatter that allow me to skip entirely ChatML parsing, waiting next llama-cpp-python release:

def fake_formatter(*args, **kwargs):
    class Dummy:
        def apply_chat_template(self, messages, **_): return messages[-1]["content"]
        def to_chat_handler(self):
            return self
    return Dummy()

You can use it like that before calling Llama(...):

import llama_cpp.llama_chat_format as chat_format
    chat_format.Jinja2ChatFormatter = fake_formatter

I see that a pull request already exists to permanently fix the problem (#2082).