What changed

This PR adds a local Codex-based translation path for PDFMathTranslate, so users can translate PDFs through the local Codex CLI without wiring up any external translation API key.

Key changes:

• add a built-in codex translator that uses the local Codex CLI
• keep the existing CLI and GUI service entry points intact
• add capability probing plus fast/compat command selection for the local CLI path
• add bounded batch translation, long-text splitting, retry, and fallback behavior so full PDFs can complete reliably
• force effective concurrency to 1 for the local CLI path instead of following the generic thread fanout
• add whitespace normalization for Chinese output so unnecessary spaces between Chinese words and punctuation are removed

Why this changed

The main goal is to make Codex usable as a local translation backend out of the box.

Before this change, users who wanted a Codex-style workflow still had to rely on external translation APIs or on a fragile short-text-only prototype. This PR makes the local CLI path practical for end-to-end PDF translation while keeping the integration surface simple.

User impact

Users can now translate with:

• local Codex CLI
• existing Codex login state
• no user-provided translation API key

Notable behavior:

• pdf2zh file.pdf -s codex is supported
• GUI keeps exposing Codex
• the local CLI translator batches text internally for better full-PDF throughput
• effective concurrency is fixed to 1 for this translator path
• Chinese output spacing is cleaned up before the translated PDF is written

Validation

Automated:

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 /opt/anaconda3/bin/python3.12 -m pytest test/test_translator.py test/test_converter.py test/test_kernel.py -q
• /opt/anaconda3/bin/python3.12 -m compileall pdf2zh test

Live checks:

• single-string smoke translation
• multi-item batch smoke translation
• long-text batch smoke translation
• full target PDF smoke run producing mono and dual outputs for MPC 跟踪控制.pdf