fix(openai): round-trip Gemini 3 thought_signature on tool calls

[OtavioTavares]

Problem

With Gemini 3 models (e.g. gemini-3.1-flash-lite, gemini-3.5-flash) via the OpenAI-compatible endpoint, multi-turn tool calling fails on the follow-up request:

400 - Function call is missing a thought_signature in functionCall parts.
This is required for tools to work correctly ...

Gemini 3 is a thinking model: each tool call carries a thought_signature that MUST be echoed back on the next request. In the OpenAI-compat protocol it is surfaced as tool_calls[].extra_content ({"google": {"thought_signature": "..."}}). openai_client.py dropped this field when parsing tool calls and never resent it, so the 2nd turn 400s.

Fix

Mirror the existing reasoning_content / _reasoning round-trip pattern for extra_content:

• Capture extra_content off each tool call (streaming and non-streaming). In streaming, Gemini sends it on a trailing index-less delta, so it is attached to the current tool call.
• Stash it on the message as _tool_extra_content (keyed by tool-call id).
• Replay it in _convert_assistant_message on the assistant tool_calls.

Gated on presence: providers that do not set extra_content (OpenAI, DashScope, GitHub Models, Kimi, ...) are unaffected.

Tests

• New TestThoughtSignatureRoundTrip (capture from attr / model_extra, replay, and omission when absent).
• Full tests/test_api/test_openai_client.py passes (46 tests).
• Verified against the live Gemini OpenAI-compat endpoint: a 2-turn tool call 400s WITHOUT the field and succeeds WITH it.