feat(gemini3) : Add Gemini 3 support with thinking_level and thought_signature

[varghesepaul]

Adds support for Gemini 3 models with model-specific parameter handling and multi-turn function calling via thought_signature.

Changes:

• Add gemini-3-pro-preview model
• Support thinking_level parameter for Gemini 3 models
• Maintain backward compatibility with thinking_budget for Gemini 2.x
• Implement thought_signature capture and injection for multi-turn function calls
• Add model version detection to route parameters correctly
• Optimize thought_signature handling to only process for Gemini 3 models
• Inject thought_signature inline during format conversion to avoid post-processing