Update benchmarks with fresh 200-doc evaluation

[jdrhyne]

Summary

Updated benchmark data in README.md and docs/benchmarks.md with fresh results from a reproducible evaluation run on 200 PDF documents with ground-truth annotations.

Changes

• Re-ran all non-ML parsers using the opendataloader-bench harness
• Added pypdf and liteparse to comparison (7 parsers total)
• Removed opendataloader-hybrid (uses docling internally, redundant)
• Added reproducibility instructions linking to the open benchmark harness

Results (200 docs, ground truth)

Solution	Overall	NID	TEDS	MHS	Speed (s/doc)
docling	0.88	0.90	0.89	0.82	0.618
Nutrient	0.88	0.92	0.66	0.81	0.007
opendataloader	0.83	0.90	0.49	0.74	0.014
pymupdf4llm	0.83	0.88	0.48	0.78	0.252
markitdown	0.59	0.84	0.27	0.00	0.106
pypdf	0.58	0.87	0.00	0.00	0.019
liteparse	0.57	0.86	0.00	0.00	0.233

Nutrient matches docling on overall accuracy (0.88) while being 90x faster.

Test plan

• Benchmark data matches opendataloader-bench evaluation.json outputs
• Speed claims match summary.json timing data
• Reproducibility link works