feat(google): add response_mime_type and response_schema params

[Panmax]

Add instance-level output format control to the Google GenAI LLM plugin. Previously, structured output could only be configured per-call via the chat() method's response_format parameter. In scenarios where the LLM is consistently used to produce JSON responses (e.g. extracting structured data, building tool-calling pipelines), having to pass response_format on every chat() call is repetitive and error-prone.

In our scenario, the agent's prompt requires the model to output JSON, which is then parsed in the tts_node to let the agent read out the content, and send an RPC to the frontend, so we need instance-level control over the JSON output.

With these two new init parameters:

• response_mime_type: controls the output MIME type (e.g. "application/json") without requiring a schema, useful for free-form JSON output.
• response_schema: constrains the JSON output to a specific structure (Pydantic model or Google SchemaUnion), automatically sets mime type to "application/json" if not explicitly provided.

The chat() response_format parameter still takes precedence when provided, preserving full backward compatibility.