Add MS MARCO large-tier benchmark path

[SonAIengine]

Summary

• add a JSONL-sharded MS MARCO passage downloader for large public-corpus evaluation
• teach the tier1 runner to load metadata JSON + corpus JSONL while preserving selected-query gold docs under corpus limits
• add optional manual MS MARCO large-tier dispatch inputs to the public scale workflow
• document the new 100k MS MARCO smoke result and ignore generated JSONL artifacts

Local large-tier result

• MS MARCO passage dev, 100k docs, 50 queries, embedder-free SQLite/EvidenceSearch
• MRR@10 0.673, R@5 0.740, R@10 0.770, Hit@10 39/50
• Build 81.9s, Search 5.4s

Tests

• uv run --extra dev ruff check examples/ablation/download_benchmarks.py examples/ablation/run_tier1_benchmarks.py tests/test_tier1_benchmarks.py tests/test_download_benchmarks.py
• uv run --extra dev ruff format --check examples/ablation/download_benchmarks.py examples/ablation/run_tier1_benchmarks.py tests/test_tier1_benchmarks.py tests/test_download_benchmarks.py
• uv run --extra dev pytest tests/test_tier1_benchmarks.py tests/test_download_benchmarks.py -q
• uv run python - <<PY
  import yaml
  from pathlib import Path
  yaml.safe_load(Path(".github/workflows/public-scale.yml").read_text())
  print("yaml ok")
  PY
• PYTHONUNBUFFERED=1 uv run --extra eval python examples/ablation/download_benchmarks.py --only msmarco_passage --large-corpus-limit 100000
• PYTHONUNBUFFERED=1 uv run --extra sqlite python examples/ablation/run_tier1_benchmarks.py --only msmarco --subset 50 --corpus-limit 100000 --use-sqlite-graph