Port chat_parsing core

[yonigozlan]

Port Transformers' standalone chat_parsing library into SGLang so checkpoint-defined response formats can be parsed without adding a Transformers runtime dependency.

Context

Chat templates convert structured messages into the token sequence expected by a model. Response templates describe the reverse operation: converting raw generated output back into structured reasoning, content, and tool calls.

Transformers implements this through transformers.utils.chat_parsing, a self-contained parser driven by tokenizer response_template metadata. The library is intentionally portable so serving projects can vendor it and integrate it with their own APIs.

What this PR adds

• Vendor the chat_parsing package from Transformers commit 9cfd6ab7c95f080dc1f33bcc7f97725ce783876e, preserving its organization and Apache license with SGLang lint and format wrappers.
• Include the declarative response-template loader, content parsers, transforms, prefix-aware response parser, and streaming region events.
• Preserve support for literal and regex delimiters, repeated fields, implicit content fields, structured content parsers, transforms, and required or optional fields.
• Require the prompt prefix when creating a response parser, matching Transformers behavior. This lets the parser understand assistant prefills or regions opened by the chat template before generation starts.
• Preserve Transformers' shallow argument coercion: strings may be converted using the declared top-level schema, while already-decoded nested values remain unchanged.
• Add registered CPU unit tests covering parser behavior and representative response-template fixtures.

SGLang-specific adaptations

• Replace Transformers-specific logging with Python's standard logging module.
• Accept the OpenAI-style tool dictionaries already used by SGLang instead of depending on Transformers tool-schema helpers.
• Add the direct regex dependency required by the parser.

This PR only introduces the generic parsing library. The richer serving event contract and SGLang adapters are kept in the following stacked PRs so each layer can be reviewed independently.

---

CI States

Latest PR Test (Base): ⏳ Run #36347873468
Latest PR Test (Extra): ❌ Run #36347873233
Latest PR Test (AMD ROCm 10): ⏳ Run #36347873400

[Jiminator]

Hey Yoni, thank for the contribution, I left a small suggestion. Feel free to fix when you have time!

[Jiminator]

LGTM

[JustinTong0323]

/tag-and-rerun-ci

[JustinTong0323]

/rerun-failed-ci

[JustinTong0323]

Pure addition, safe to merge.