Add pdf_oxide to benchmark suite

[yfedoseev]

Adds pdf_oxide to the benchmark suite with text extraction and image extraction support.

Changes

• pdf_benchmark/library_code.py: Added pdf_oxide_get_text() and pdf_oxide_image_extraction() functions (uses tempfile approach, same as pdftotext, since pdf_oxide accepts file paths)
• benchmark.py: Registered pdf_oxide as a library with imports
• requirements/main.in: Added pdf-oxide dependency

About pdf_oxide

• Rust core with Python bindings via PyO3
• MIT / Apache-2.0 licensed
• Text extraction, image extraction, markdown conversion
• v0.3.6 released: PyPI | GitHub

[bosd]

@yfedoseev This tool looks very promising.
Love the speed. But it does seem to struggle with text directions.
Hope the accuracy can go up from 80% to something in the 90% 's

[yfedoseev]

@bosd Thanks for the feedback! Accuracy has been a major focus, and the text-direction / reading-order handling has improved significantly over the last several releases.

I've bumped this PR to pdf_oxide 0.3.57. Re-running the benchmark's own scoring (Levenshtein ratio vs the ground-truth files) across all 14 documents:

Metric	0.3.6 (original)	0.3.57
Median	64.4%	92.7%
Mean	67.0%	89.7%

So the typical document is now solidly in the 90's, up from the ~80% you saw earlier. The mean is held back by a couple of harder layouts that we're still improving — and more accuracy gains are landing over the coming weeks. A re-run on your side would be very welcome to confirm.