Honor source splits throughout benchmark runs

[davanstrien]

Summary

• Forward --split to every OCR job, including dry runs and the default train split.
• Record the source split in evaluation metadata and dataset cards.
• Make the viewer load images from the recorded split and use the latest append-only metadata row.
• Keep old results compatible by defaulting missing or null split metadata to train.

Why

Non-train runs could judge one split while jobs or the viewer silently read train, producing incorrect or broken results.

Validation

• 615 tests passed
• Ruff passed on src and tests
• ty passed on src
• git diff --check passed