Chat template rendering extensions to match transformers

[CISC]

I've been using the jinja2 tojson filter in a couple of chat templates to render function calling instructions, which is working very well, however sometimes this will introduce escaped unicode characters, which is undesirable.

This PR ~~simply changes the default parameters of tojson~~ replaces tojson so that the JSON is rendered in unicode without escaped HTML characters.

Examples of models using tojson:

• openfunctions-v2
• Mistral-7B-Instruct-v0.3

I've also removed the leftover (and unnecessary) loader parameter since I was touching this code area anyway.

Update:

Further changes to chat templates are being made in transformers and since they are dependant on the changes already made here I will simply append them to this PR:

• Added loopcontrols jinja2 extension (adds break and continue)
• Added strftime_now to render the current time

See huggingface/transformers#32684

[CISC]

Since the merging of a similar patch to transformers and the formalizing of tools (see references in #1336) there are starting to pop up a lot more models using tojson in their chat templates.

[CISC]

@abetlen Llama 3.1 is the latest model to use tojson, would be nice to have matching behaviour as transformers.

[CISC]

@abetlen Even more transformers changes incoming, see updated OP.

[CISC]

@abetlen It would be really nice to have this merged soon, it's starting to cause issues with quite a few templates not to have this in place.

Additionally this is a requirement for implementing inverse templates (which I'd like to do, as soon as it's merged in transformers, since it will be super useful for universal function calling and parsing).

[serhii-nakon]

Hello I support to merge it. I have issues with Mistral models (devstral too)

[CISC]

@abetlen I see strftime_now support was finally added, unfortunately that is not enough, PTAL.