Add portable prediction and label artifacts

[adrian-prior]

Generated by Codex

Add portable prediction artifacts and shared label artifacts for offline evaluation. Store config identity, ordered rows, split metadata, metrics and fit/predict timings alongside predictions. Validate alignment and artifact integrity when loading, with labels deduplicated by content.

Scope and validation

• Adds the artifact writer and public loading functions. Runner and CLI integration is in the next stack layer.
• Reuses existing run identity rather than recomputing input-table checksums.
• Local artifact I/O only. Existing prediction paths are not overwritten; this is not a resumable prediction cache.
• Lower layer: 38 focused tests and all pre-commit hooks pass.
• Complete stack: 440 tests pass, 11 skip; pinned Ruff checks, distribution build and required-package-data verification pass on Python 3.11.