common : replace wrap_for_generation with a prefix convenience function + gpt-oss fixes

[aldehir]

Overview

Remove wrap_for_generation function and replace with a prefix() parser that generates a literal up to a given delimiter.

Clean up gpt-oss and fix required tool calling. Injecting the generation prompt greatly simplifies the gpt-oss parser rules. This assumes <|start|>assistant is always a prefix of generation_prompt, which is a reasonable assumption. It will work even if users append extra tokens, i.e. <|start|>assistant<|channel|>analysis<|message|>I am thinking of chocolate<|end|>.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: No

cc @pwilkin

[pwilkin]

Dunno, the creator of the PEG parser said that it's not a good idea to add more helper builders ;)

[aldehir]

I meant inside the core peg parser so it's still useful for general parsing tasks. Inside the chat builder, anything goes!

Sorry for not making that distinction. I do prefer more composable patterns though, for example we're just matching a prefix here.

[pwilkin]

Aight, might want to take this out of draft :)

[aldehir]

Wanted to run through a few models before removing out of draft, ministral-3 thinks forever.

[pwilkin]

BTW you might want to take a look at #20844 - I scrapped the previous approach and did the regexp generator for the full pattern, but also expanded the tests to check whether the grammar is actually triggered.