feat: replace LLM provider with Sakana Fugu (straight swap)

From the PR description

Summary

Replaces the existing LLM provider with Sakana AI's Fugu model - a straight permanent swap, no feature flag.

Changes

• backend/src/lib/llm/sakana.ts - Sakana Fugu adapter (OpenAI Chat Completions-compatible, streaming SSE)
• backend/src/lib/llm/{types,models,index}.ts - provider abstraction layer; completeText and streaming route to Sakana
• backend/src/routes/chat.ts - stamps provider_metadata: { provider_name, model_name } on every assistant chat_messages insert (success, abort, and error paths) via SAKANA_PROVIDER_METADATA constant
• backend/src/index.ts - startup guard: server refuses to boot without SAKANA_API_KEY
• supabase/migrations/20260623000000_add_provider_metadata.sql - adds provider_metadata JSONB column to chat_messages
• backend/src/lib/llm/__tests__/sakana.test.ts - unit tests for the adapter
• backend/.env.example and backend/README.md updated

Deployment checklist

1. Run the migration before deploying the new backend
2. Set SAKANA_API_KEY in Railway (Settings → Variables)
3. Optionally set SAKANA_MODEL (defaults to fugu-ultra-20260615)
4. Remove any legacy FUGU_API_KEY or LLM_PROVIDER env vars after merge

Cost note

Sakana Fugu is priced at $5 / M input tokens and $30 / M output tokens.