[codex] Support localized ChatGPT UI and new intelligence picker

[LucasYeh702]

Summary

• Add localized ChatGPT UI label fallbacks for Traditional Chinese controls.
• Support the newer composer intelligence picker without requiring the old Configure... menu item.
• Parse localized and compact visible intelligence labels such as Pro 延伸, GPT-5.5, and Chinese effort labels.
• Detect localized response completion actions such as 複製回應 and 更多動作.

Why

ChatGPT's web UI can be localized, and the Pro model selector has moved from the older Configure flow to a composer intelligence picker. In Traditional Chinese UI, GPT Relay could fail to find the composer, model selector, or completion controls even when the requested GPT-5.5 Pro Extended selection was visible.

Validation

• /Users/liwei/.cache/codex-runtimes/codex-primary-runtime/dependencies/node/bin/node --check plugins/gpt-relay/scripts/chatgpt_relay.mjs
• /Users/liwei/.cache/codex-runtimes/codex-primary-runtime/dependencies/node/bin/node --test plugins/gpt-relay/scripts/chatgpt_relay.test.mjs
• Live smoke test with ChatGPT Traditional Chinese UI: 5.5 Pro Extended selected as Pro 延伸, prompt returned READY.

[LucasYeh702]

Maintainer note: I opened this as a localization-resilience fix rather than only a Traditional Chinese patch.

The immediate breakage I could reproduce was in a zh-Hant ChatGPT UI, where the composer/model controls show labels such as 與 ChatGPT 對話, Pro 延伸, GPT-5.5, 複製回應, and 更多動作. The broader suggestion is to treat ChatGPT web UI labels as localized data and keep the selector logic tolerant of locale-specific wording and layout changes.

This PR starts with the locale I could verify live, but the same pattern should be extendable to other languages by adding label alternatives and regression tests instead of hard-coding English-only selectors. If you prefer a different structure, such as per-locale dictionaries or a more generic selector strategy, I am happy to adjust the patch in that direction.