data: build v0.9 cross-benchmark passfail packs

[AbdelStark]

Summary

• Add multi-source pass/fail pack inputs for HumanEval + MBPP-Plus style completion-label artifacts while keeping legacy single-source mode.
• Split by benchmark/problem, add optional held-out split coverage gates, and fail typed split_coverage_blocker errors when required labels are not evaluable.
• Enrich pack reports/manifests with benchmark counts, output-magnitude coverage, held-out coverage, readiness gates, and explicit benchmark_id row metadata.
• Document the v0.9 multi-source script mode and data-model row contract.

Validation

• uv run pytest tests/data/execution_pack/test_passfail_pack.py
• uv run pytest tests/data/execution_pack tests/training/test_execution_pack_loader.py tests/docs/test_usage_docs.py tests/docs/test_execution_substrate_docs.py
• uv run python -m compileall -q -x 'tests/fixtures/codestate/invalid_(before|after)\.py$' codelewm tests scripts/build-passfail-pack
• uv run scripts/build-passfail-pack --help
• git diff --check
• uv run pytest tests/ (941 passed, 8 skipped, 1 warning)

Closes #387