Export model predictions from benchmark runs

[adrian-prior]

Generated by Codex

Export model validation and test predictions with --predictions-dir so results can be evaluated offline. Add --refit-all-configs to generate test artifacts for successful configs beyond the selected/default configs. Keep extra fit/predict timings in artifact metadata, separate from benchmark result timings.

Behavior and validation

• Normal export mode reuses tuning predictions and selected/default test predictions without adding fits.
• Separate benchmark and extra-refit loops make the timing policy explicit.
• Shared validation labels follow the evaluated inner rows; test labels are exported after prediction.
• Extra timings cover fit and predict calls, not scoring or artifact I/O.
• Tests verify classification and regression round trips, metric reproduction from saved labels, row/class/split mismatch rejection, label corruption detection, fit counts and timing separation.
• Full stack: 440 tests pass, 11 skip. All pre-commit hooks, pinned Ruff checks, distribution build and required-package-data verification pass on Python 3.11.