Parallelize OCR block processing and add optional CUDA preprocessing

[zeetee1235]

Motivation

• Reduce wall-clock OCR latency without changing fusion or parser decision logic by running independent per-block Tesseract calls concurrently.
• Provide an optional fast preprocessing path that can leverage OpenCV CUDA when available to accelerate grayscale/threshold operations.

Description

• Added a configurable worker pool to ocr/bridge/ocr_bridge.py to process OCR blocks in parallel using ThreadPoolExecutor and DOCSTRUCT_OCR_WORKERS to control concurrency while preserving deterministic output ordering.
• Introduced CUDA-aware preprocessing helpers _to_grayscale and _otsu_threshold that use OpenCV CUDA if DOCSTRUCT_OCR_USE_CUDA=1 is set and CUDA bindings are available, otherwise silently fall back to CPU implementations.
• Kept the original multi-pass OCR candidate scoring, block classification, post-processing, and LaTeX (pix2tex) handling intact and re-applied them to the ordered results to preserve existing OCR logic.
• Documented the new runtime environment variables in README.md (DOCSTRUCT_OCR_WORKERS, DOCSTRUCT_OCR_USE_CUDA, and existing DOCSTRUCT_BRIDGE/DOCSTRUCT_PYTHON).

Testing

• Ran Python syntax check with python3 -m py_compile ocr/bridge/ocr_bridge.py, which completed successfully.
• Ran Rust test suite with cargo test -q, and the tests completed successfully (all automated tests passed).

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: df510233a3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid Python 3.10-only union syntax in bridge

The new annotation -> dict | None in run_ocr uses PEP 604 syntax, which is only valid on Python 3.10+, so the bridge now fails to import with a SyntaxError on Python 3.9 runtimes that previously worked; this is a functional regression because the OCR process cannot start at all in those environments (README still documents Python 3.8+ support).

Useful? React with 👍 / 👎.