feat: update Gemini model support with latest Google models

[adryserage]

Summary

Google has released several new Gemini models that should be supported in Spacebot. The current implementation only uses gemini-2.5-flash for all process types with no fallback chain.

New Models to Add

Model ID	Type	Key Capabilities
gemini-3.1-pro-preview	Preview	Most advanced intelligence, agentic coding
gemini-3-pro-preview	Preview	Advanced multimodal understanding with reasoning
gemini-3-flash-preview	Preview	High performance rivaling larger models at lower cost
gemini-2.5-pro	Stable	Best for complex tasks, deep reasoning, coding
gemini-2.5-flash	Stable	Already supported - fast reasoning
gemini-2.5-flash-lite	Stable	Fastest, most economical multimodal option

Changes Required

1. Routing Defaults (src/llm/routing.rs)

• Use gemini-2.5-pro for channel/branch (quality tasks)
• Use gemini-2.5-flash for worker/compactor/cortex (speed tasks)
• Add fallback chain: gemini-2.5-pro → gemini-2.5-flash → gemini-2.5-flash-lite

2. Voice Transcription Models (src/api/models.rs)

• Add native Gemini API model paths (currently only OpenRouter paths exist)
• Add all models that support input_audio via the native Gemini endpoint

3. Model Catalog

• The models.dev catalog handles most discovery automatically
• Verify mapping for new Gemini 3.x preview models

Reference

• Gemini API Models Documentation

[adryserage]

Addressed in PR #134