fix: retry transient OpenAI Codex request failures

[guanzi008]

Summary

• retry transient OpenAI Codex backend failures instead of failing on the first malformed/incomplete response
• retry transport-level request errors up to five attempts while keeping token refresh single-shot
• add coverage for transient backend failures, transport timeouts, and retry exhaustion

Testing

• .venv/bin/pytest -q tests/test_openai_codex.py