Add resumable model-version semantic backfills

[haasonsaas]

Context

The schema stores extractor/model versions, but there is no operator-facing workflow for extracting a new semantic version across previously retained artifacts. Reprocessing a large tenant should not require a single unbounded request.

Proposal

Add a corpusctl backfill command and worker workflow that selects artifacts by tenant, format, architecture, and model version; supports dry-run, bounded batches, resume tokens, and cancellation; and reports progress.

Acceptance criteria

• Backfills never compare incompatible formats or architectures.
• A restarted worker resumes without duplicating feature rows or edges.
• Operators can preview counts and estimated work before enqueueing.
• Progress and terminal failures are queryable by tenant and model version.

References

• versioned feature schema
• semantic persistence
• CLI crate