Add Google Gemini 2.5 Pro as a provider for the Inference effect. Unlike the OpenAI-compatible providers (Mistral, xAI, DeepSeek), Gemini uses a distinct API shape that requires custom request/response handling — similar to how Anthropic is handled today.

API details

• Base URL: https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent (model embedded in URL)
• Auth: query param ?key={API_KEY} or header x-goog-api-key: {key}
• Default model: gemini-2.5-pro
• Request body: uses contents array with {"role": "user", "parts": [{"text": prompt}]} instead of messages
• Response path: data["candidates"][0]["content"]["parts"][0]["text"]
• Env var: VERA_GEMINI_API_KEY

Implementation notes
The request body format and response path differ from both the OpenAI-compatible providers and Anthropic, so this cannot reuse the shared OpenAI dispatch path. After #413 lands with the _ProviderConfig registry, Gemini will need either:

1. Custom fields in _ProviderConfig for request_format and response_path, or
2. A small _gemini_call helper alongside _anthropic_call (Anthropic's current pattern)

Option 2 is simpler and consistent with the existing Anthropic special-case. The _PROVIDERS registry can dispatch to a callable instead of a config when custom logic is required.

Depends on #413.