Implement the im_start / im_stop funkiness

[lukehinds]

[jhrozek]

I actually realized we need to do this for vllm as well, while it does use the openai message format it also uses the im_start/im_end tags. We side-stepped the problem in the llama.cpp provider to not use them, but not the vllm provider.

[lukehinds]

How are we on this @jhrozek ?

[lukehinds]

SHould be ready to merge @jhrozek will comment out function for now, in case we need it, but until then the provider routing should take care fo this.

[jhrozek]

This was finally merged