perf: coalesce concurrent document normalization

[dajiaohuang]

Summary

• coalesce concurrent normalized-document cache misses by immutable content/profile target
• recheck the cache after waiting so one caller performs the PDF/OCR pipeline and followers reuse its result
• retain parallel conversion for distinct cache keys and avoid retaining an unbounded lock registry

RapidOCR inference remains behind its existing shared per-model lock; this change removes duplicate same-document work without assuming the OCR or PDF runtime is safe to parallelize.

Verification

• pytest (310 passed, 14 skipped)
• ruff check .
• uv build --wheel
• imported normalize_document from the built wheel
• git diff --check