Benchmark current parser against Docling and MinerU

[EstaraAtopos]

Summary

• compare the current parser, Docling 2.117.0, and MinerU 3.4.4 on 11 synthetic English/Chinese PDFs
• reuse the existing KnowledgeChunk builder and fixed multilingual retrieval backend
• select Docling only as a conditional adapter candidate for scanned and simple-table PDFs
• keep multi-column, mixed-layout table, and formula cases excluded; production routing remains unchanged

Verification

• related parser/layout/chunk tests: 66 passed, 1 skipped
• full pytest: 1722 passed, 5 skipped
• dev_check: 1720 passed, 7 skipped; migration and API smoke passed
• release safety: passed
• external API requests: 0
• real Provider requests: 0
• accident database unchanged

Important boundary

This PR is evaluation-only. It does not connect Docling to production ingestion, change KnowledgeChunk or retrieval semantics, modify the PDF reader, or redesign concept cards.