refactor: move subagents to google gemini-3.5-flash-lite and fix google baseURL (issue #234)

[nmwael]

WHY

Issue #234 investigation proved groq unviable (free-tier TPM too low for our prompts, models reject tool calling) and found the google provider baseURL was broken (opencode's built-in google provider uses the native Gemini API, not the /v1beta/openai openai-compat endpoint). All subagents were assigned to nvidia/groq models that cannot run the full role prompts.

WHAT

• provider.google: baseURL -> https://generativelanguage.googleapis.com/v1beta, removed 'api: openai' and stale gemini-1.5-pro-latest model map; removed duplicate groq provider block (original file had two 'groq' keys - invalid JSON)\n- All 5 agents (architect, developer, tester, security-auditor, ux-ui) -> model google/gemini-3.5-flash-lite (verified working on full role prompts)\n- architect.prompt now describes the architect as a subagent orchestrator that delegates to developer/tester/security-auditor/ux-ui subagents via the Task tool\n- New test: test/unit/opencode_config_validation_test.dart validates the config requirements

HOW

Gemini 3.5 Flash Lite and 3.1 Flash Lite were the only models verified to handle the full architect prompt (~12.7k tokens) and real role tasks end-to-end in opencode. Config is now validated by an executable test that reads opencode.json. Small_model, main model, and other providers untouched.