feat: switch to threaded PDF pipeline (docling 2.99)

[selloriwoo]

Summary

Switch the docling PDF converter to docling's threaded pipeline (ThreadedStandardPdfPipeline + ThreadedPdfPipelineOptions), which processes pages through concurrent stages (parse / layout / table) connected by bounded queues. Multi-page PDFs convert faster than the sequential StandardPdfPipeline.

The threaded backend requires docling-parse v6 (PR docling-project/docling#3377, landed in docling 2.96), so this also bumps docling.

Changes

• python/run_docling.py: build_converter now wires ThreadedStandardPdfPipeline via PdfFormatOption(pipeline_cls=...) and builds ThreadedPdfPipelineOptions. Option fields and defaults are unchanged.
• python/pyproject.toml: docling>=2.91.0 → >=2.99.0.
• python/uv.lock: docling 2.99.0, docling-parse 6.2.0 (+ dependency restructure into docling-slim; macOS-only ocrmac/pyobjc dropped in favour of rapidocr, matching docling 2.99 defaults).

Test

• cargo test -p agent-k --features internal translate_pdf_from_financebench -- --ignored passes (283s): builds
  the PyInstaller bundle with docling-parse v6 and converts a real
  FinanceBench PDF end-to-end through the Rust → bundle path.
• Threaded pipeline confirmed active (pipeline_cls = ThreadedStandardPdfPipeline).

[nuri-yoo]

LGTM

[grf53]

Before merge, it seems essential to not only pass the translate_pdf_from_financebench test but also verify whether the same results (or results without quality degradation) are obtained based on existing Docling standards for several PDF documents (sampling from those at https://github.com/brekkylab/knowledge-base-examples?).

Additionally and optionally, approximate range of performance improvement?

[selloriwoo]

Spec: Ryzen 5600X, RAM: 32GB, NVIDIA RTX 3070 (8 GiB)

PDF	Pages	2.91 (s)	2.99 (s)	Speedup
10K	158	125.28	113.29	1.11×
10K	165	131.75	111.38	1.18×
10K	250	120.54	80.72	1.49×
10K	346	157.94	114.78	1.38×
10K	107	85.58	63.95	1.34×

This is the difference between ver 2.91 (baseline) and 2.99 (multithread on), based on 5 PDFs.

Our experiments show no difference in output between the single-threaded(2.91) and multithreaded(2.99) runs.

[grf53]

Thank you for your confirmations!