This PR addresses several critical issues with the Google OpenAI-compatible Tier, specifically
targeting Gemini 3 models which require strict "thought signature" handling and exhibit
non-standard termination behaviors.

Key Changes:

1. Forced tool_use Stop Reason:

   • Gemini 3 models frequently return finish_reason: "stop" even when tool calls are
     present in the response.
   • Updated build_stream_response and the OpenAI stream loop to force stop_reason:
     "tool_use" whenever the tool call map is non-empty. This ensures the agent_engine
     correctly triggers the execution turn.

2. Corrected Thought Signature Handling:

   • Parsing: Updated the response parser to capture thought_signature from
     extra_content.google (where Google places it) rather than incorrectly looking inside
     the function object.
   • Propagation: Updated translate_messages_to_oai_with_reasoning to send the signature
     back in the subsequent turn using the exact nested structure Google requires:
     tool_calls[].extra_content.google.thought_signature. Failing to do this results in 400
     Bad Request errors on Turn 2.

3. Improved Stream Chunk Indexing:

   • Google’s API sometimes delivers tool call deltas without an index field.
   • Updated process_openai_stream_event to handle missing indexes by matching on the tool
     id or assigning a new unique index. This prevents tool call chunks from overwriting
     each other and ensures all signatures are captured.

4. Serialization Stability:

   • Updated OaiToolCall and related structs to properly deserialize extra_content for
     non-streaming responses.
   • Ensured thought_signature is correctly propagated through ResponseContentBlock and
     ContentBlock types.

Verified using an automated test suite across multiple scenarios that did not work with #211

• Non-thinking queries: Basic chat works correctly.
• Thinking queries: Reasoning content is successfully captured and wrapped in tags.
• Multi-turn Tool Use: Verified Turn 1 (Thought + Tool Call) followed by Turn 2 (Tool Result
  • Final Answer) works without 400 errors or premature stops.