[zchryst]

Summary

Fixes three Gemini 3 / 3.5 thinking, context caching, and usage metadata issues in GeminiProvider (aisuite/providers/gemini_provider.py) and GoogleProvider (aisuite/providers/google_provider.py):

1. ThinkingConfig Routing, cached_content Passthrough, and Thought Separation:
   • Added "thinking_config" and "cached_content" to CONFIG_KEYS in aisuite/providers/gemini_provider.py.
   • Added GeminiProvider._build_thinking_config() to translate reasoning_effort, thinking, thinking_level, thinking_budget, and include_thoughts kwargs into google.genai.types.ThinkingConfig, routing thinking_level for Gemini 3+ Flash/Pro (with "minimal" on Flash and "low" minimum on Pro) and thinking_budget for Gemini 2.5 (with regional prefix normalization such as au.gemini-3.5-flash).
   • Updated GeminiMessageConverter.parse_candidate() and convert_response() so thought parts (part.thought == True) are separated into Message.reasoning_content instead of being concatenated into Message.content.
   • Updated DEFAULT_TEMPERATURE in aisuite/providers/google_provider.py to None instead of forcing 0.7 on Gemini 3+ models.
2. Usage Metadata (thoughts_token_count / reasoning_tokens):
   • Updated GeminiMessageConverter.get_completion_usage() to extract thoughts_token_count into CompletionTokensDetails(reasoning_tokens=...) and include (thoughts or 0) in the fallback total_tokens calculation.

Testing

• Added unit tests in tests/providers/test_gemini_provider.py and tests/providers/test_google_provider.py covering thinking_config and cached_content routing across Gemini 3.5 Flash/Pro (including regional prefixes) and Gemini 2.5, thought part separation into reasoning_content, thoughts_token_count usage details, and default temperature.
• Verified formatting with black --check and ran pytest (36 passed).