Add `anything` and `until` method in the DSL for reasoning models

[rlouf]

This was triggered by trying to make structured generation with DeepSeek, where we want to let the model generate anything (unstructured) between the think tags, and start structured generation after </think> has been generated. It would thus make sense to have an anything construct that corresponds to .*, the unstructured case. Intuitively, we could add until method so that users could write the following for a classification task

from outlines.types import anything, either

model = ...

outline = "<think>" + anything.until("</think>") + "</think>" + either("yes", "no")
result = model.generate("Are you a reasoning model?", outline)

The implementation is trickier than it looks, anything.until("<think>") is best expressed with a negative lookahead:

((?!<\/think>).)*

but the regex engine that outlines-core uses does not support lookaheads. However, it is equivalent to the following regular expression:

([^<]|<[^\/]|<\/[^t]|<\/t[^h]|<\/th[^i]|<\/thi[^n]|<\/thin[^k]|<\/think[^>])*

The idea would thus be to have anything.until generate a Regex node with an "expanded lookahead". Computing the expansion shouldn't be too hard.

Note: we could also decide that "until" keeps the </think> token. We should discuss.

Regexes

It is tempting to also implement until_regex this way. It is very simple for simple patterns like [0-9]:

(^[0-9])*

however it can get complicated quite quickly, for instance anything.until_regex("(abc|def)") which we would need to translate to:

([^ad]|a[^b]|ab[^c]|d[^e]|de[^f])*

so we will not implement this in the first iteration.