modellist: automatically replace known chat templates with our versions

[cebtenzzre]

This is a temporary fix for issues such as #3309, #3282, and #3263 until we can improve Jinja2Cpp itself.

The workaround for users loading a GGUF they downloaded from HuggingFace and then getting an ugly chat template that is in many cases incompatible with Jinja2Cpp, is to make a list of known templates and automatically replace them when found.

This is currently done silently and transparently, aside from a log message that looks like this:

[Warning] (Thu Dec 19 14:25:09 2024): automatically substituting chat template for "SummLlama3.2-3B-Q4_0.gguf"

We can make this list less necessary by improving Jinja2Cpp (see the comments in jinja_replacements.cpp for next steps), but for now this is better than what the current release does with these models.

This allows us to be compatible with e.g. any finetune of Llama 3.2 3B Instruct that does not alter the chat template, without having to touch models3.json.

---

models3.json has been touched in this PR for consistency with the list of substitutions, which now covers all official GPT4All models, even the ones that were working before. This makes them prettier and easier to edit.