Support Castorini prebuilt indexes, topics, and qrels

[lilyjge]

Summary

• load the live Castorini prebuilt-index and Anserini topics/qrels catalogs instead of maintaining a hard-coded asset list
• add piika prebuilt discovery and setup commands with topic normalization, checksum verification, atomic downloads, mirror fallback, and safe archive extraction
• register installed benchmark manifests dynamically and preserve upstream source provenance in run snapshots
• document the workflow, including MS MARCO V2 Passage with DL21

Validation

• npm run check
• npx tsx --test tests/prebuilt_catalog.test.ts tests/prebuilt_setup.test.ts tests/benchmarks/registry.test.ts (19/19 passing)
• live catalog dry-run for msmarco-v2-passage + dl21 + dl21-passage
• real two-query smoke run against the installed MS MARCO V2 Passage index using the Pyserini REST two-tool interface, openai-codex/gpt-5.5, and medium thinking; zero tool failures

The complete npm test suite was also run. Its fixture-dependent tests fail in this checkout because ignored benchmark assets are absent under data/benchmark-template, data/browsecomp-plus, and data/msmarco-v1-passage; the scoped and static checks above pass.