fix(chat-completions): normalize provider thinking aliases

[stemirkhan]

Summary

• normalize provider-specific thinking / enable_thinking aliases into the canonical reasoning payload
• apply the normalization in Chat Completions request mapping and in the shared OpenAI-compatible Responses sanitation path
• add unit + integration coverage and an OpenSpec change for the compatibility behavior

Why

Fixes #205.

Some OpenAI-compatible clients reuse provider-specific reasoning controls when switching their base_url to codex-lb. Those aliases were leaking through to the upstream ChatGPT backend unchanged, which caused avoidable upstream validation failures instead of a smooth compatibility path.

This keeps the fix intentionally small:

• enable_thinking: true maps to reasoning.effort = "medium"
• Anthropic-style thinking payloads are normalized to the same canonical shape
• explicit reasoning still wins over provider aliases

Testing

• /home/temirkhan/codex-lb-my/codex-lb-upstream/.venv/bin/python -m pytest tests/unit/test_chat_request_mapping.py tests/unit/test_openai_requests.py tests/integration/test_openai_compat_features.py -q
• openspec validate --specs

Notes

• scoped to OpenAI-compatible request compatibility only
• no API surface expansion and no dashboard changes