wip(gemini 3): Add support for streaming Gemini 3 thoughts

[ksheasby-va]

This pull request enhances the handling of "thinking" (internal reasoning) content for Google Gemini 3 models in the llms/googleaiv2 integration. The changes ensure that thought content is extracted, streamed, and returned in a standardized way, improving compatibility with other providers and enabling richer reasoning features. Unit tests are also updated to verify the new behavior.

Gemini 3 "thinking" content extraction and streaming:

• Refactored the convertThinkingConfig function to only set ThinkingLevel for HIGH mode, enabling more natural "thinking" for other modes by just enabling IncludeThoughts without forcing extended behavior.
• Updated convertCandidates to extract and separate reasoning/thought content from regular text, storing it in both the response and metadata (ReasoningContent and ThinkingContent). [1] [2] [3]
• Enhanced streaming support to emit thought chunks via a new StreamingReasoningFunc callback, prevent duplicate thoughts, and accumulate all streamed thoughts for inclusion in the final response. [1] [2] [3] [4] [5] [6]

Testing improvements:

• Added and extended unit tests to verify extraction and correct assignment of thinking content, including checks for content separation, metadata, and token counts. [1] [2]

DALE-937