Add support for OpenAI's `n` parameter to generate multiple candidates:
- OpenAI→Gemini: Map `n` to `generationConfig.candidateCount` when > 1
- Gemini→OpenAI: Map `candidateCount` back to `n` parameter
- Response: Refactor to iterate over all candidates and return multiple
choices, properly tracking function call indices per candidate
This enables requesting multiple completions per request when using
Gemini models through the OpenAI-compatible API.
Ported from upstream commits b0c17af2, 14c74e5e, 5bb9c2a2.