feat: add json_object support in response_format

[windreamer]

Motivation

OpenAI-compatible response_format={"type": "json_object"} is still part of the official spec and is explicitly requested by users.
lmdeploy currently rejects this value with
ValueError: unsupported format type: json_object.
We need to accept it and force the model to emit any valid JSON object without requiring the caller to supply a schema.

Modification

1. GuidedDecodingManager

   • Treat json_object as a built-in schema:
     '{"type": "object", "additionalProperties": true}'
   • Dispatch it to the existing JSON-schema processor.

2. TurboMind backend

   • Add json_object branch when extracting the schema from response_format.
   • Compile the same permissive object schema with the grammar compiler.

3. OpenAI-API server

   • Remove the outdated TODO comment; no extra logic is needed here after the core change.

4. Tests

   • Extend the guided-decoding matrix to cover json_object, json_schema, regex_schema, and the un-guided case.
   • Validate json_object responses with the same permissive schema used in the engine.

With this patch, clients can use
response_format={"type": "json_object"} and receive guaranteed valid JSON objects.

[grimoire]

LGTM

[jingyibo123]

@windreamer How does this relate to the usage of openai.client.beta.chat.completions.parse along with pydantic model for structured output?
Will that work out of the box after this PR, or does that require additional adaptation?

[windreamer]

openai.client.beta.chat.completions.parse

I haven't dug into the implementation details of OpenAI SDK. But from my point of view, the parse API is just a grammar-sugar on top of structure output (aka guided decoding). This should be working out of box with this PR.