fix: use config API key for embeddings

[24601]

Summary

• resolve the embedding client API key from either OPENAI_API_KEY or ~/.gbrain/config.json
• rebuild the OpenAI client when the resolved key changes so tests can cover env/config precedence cleanly
• add a focused regression test for config-backed embedding credentials

Why this matters

gbrain embed --stale currently assumes OPENAI_API_KEY is present in the shell environment. In real local installs, gbrain init persists openai_api_key in ~/.gbrain/config.json, and other commands already rely on config loading. That leaves embeddings in a broken state for users whose shell env is clean but whose saved config is valid.

Test plan

• bun test test/embedding-config.test.ts test/embed.test.ts
• bun test
• bun run test:e2e (blocked locally: no reachable test Postgres at localhost:5434 in this environment, ideally CI checks will spin up/set up an env appropriate for this)