feat(llm): support OpenAI Codex endpoints

[linux-devil]

Summary

• add an OpenAI Responses API client selected by /v1/responses endpoints
• send max_completion_tokens for GPT-5/Codex/o-series chat-completions models while preserving max_tokens for legacy models
• make OCR environment variables take precedence over config and accept CI-compatible OCR_LLM_AUTH_TOKEN / OCR_LLM_USE_ANTHROPIC names
• document Codex/GPT-5 setup examples

Tests

• git diff --check
• go test ./internal/llm
• make build
• local ocr llm test against an Azure GPT-5.5 chat-completions deployment

Note: go test ./... currently fails in internal/config/rules because existing tests expect Chinese default-rule snippets while the embedded default rules resolve to English text; that failure is unrelated to this LLM endpoint change.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[github-actions]

🔍 OpenCodeReview found 8 issue(s) in this PR.

• ✅ 8 posted as inline comment(s)
• 📝 0 posted as summary (missing line info)

---

📊 Posting Statistics:

• ✅ Successfully posted: 6 comment(s)
• ❌ Failed to post: 2 comment(s)

❌ Failed Comments Details

• internal/llm/resolver.go: Unprocessable Entity: "Line could not be resolved"
• internal/llm/resolver.go: Unprocessable Entity: "Line could not be resolved"

[vanducng]

exactly what im looking for. hope this will be resolved and merged soon