fix(llm): use non-streaming Gemini Pro requests

[0xCheetah1]

Summary

• Force Gemini Pro reasoning models through the non-streaming /v1/messages path so their thinking budget is preserved by the gateway.
• Convert non-streaming Gemini responses back into Franklin's internal stream chunks so the rest of the agent loop remains unchanged.

Verification

• npm run build
• node dist/index.js --model google/gemini-2.5-pro --prompt "say ok"
• node dist/index.js --model google/gemini-3.1-pro --prompt "say ok"
• node dist/index.js --resume session-2026-05-07T10-51-11-210Z-fe4ab9e0 --prompt "hey"