Improve PyMuPDF4LLM OCR pipeline and benchmark results

[hashiromer]

Summary

Improves the existing pymupdf4llm_markdown pipeline and enables layout and table evaluation.

This PR also updates the pipeline documentation, dependency setup, tests, and leaderboard results. No additional PyMuPDF4LLM pipeline variants are added.

Benchmark Results

The full ParseBench run completed successfully:

• Total examples: 2,078
• Successful: 2,078
• Failed: 0

Metric	Score
Overall	50.27
Tables	57.05
Charts	2.29
Content Faithfulness	75.25
Semantic Formatting	58.39
Visual Grounding	58.38

Reproducing the Results

The submitted results were produced in the following environment:

• CPU: 8 Intel Haswell-class vCPUs
• Memory: 24 GiB RAM
• Operating system: Ubuntu 26.04 LTS
• Python: 3.12
• Tesseract: 5.5.0
• Maximum concurrency: 1
• Full-run time: approximately 1 hour

Install Tesseract with English language data:

sudo apt-get install tesseract-ocr tesseract-ocr-eng

Install the pipeline dependencies:

uv sync --python 3.12 --extra pymupdf4llm

Run the small test benchmark:

uv run --python 3.12 --extra pymupdf4llm parse-bench run pymupdf4llm_markdown --test --max_concurrent 1

Run the full benchmark used to produce the submitted scores:

uv run --python 3.12 --extra pymupdf4llm parse-bench run pymupdf4llm_markdown --max_concurrent 1

Validation

• Full benchmark completed without inference failures
• 297 tests passed
• Ruff and mypy checks passed
• Lockfile validation passed