Fix loading saved evaluation results and script exit codes

[ishantamrakar]

For items 1 and 2 of #23; item 3 is left open there.

• SparseEvalResult.load_from_npy rebuilds ControlPointSummary objects from the dicts written by save_as_npy, so the metric functions work on a reloaded result.
• evaluate_wrt_control_points.py, evaluate_wrt_pgt.py and evaluate_wrt_mps.py exit with status 1 when run() returns False.

Tested on sequence_1_19 from the demo, with the pinned pycolmap/pyceres on Linux:

• Scores and recalls printed by the scripts are unchanged (CP score 49.8639 for the demo OpenVINS estimate, 69.9432 for the pGT; evaluate_wrt_pgt.py recalls unchanged).
• A result saved by evaluate_wrt_control_points.py and reloaded with load_from_npy gives the same score and recall as the in-memory result.
• Graceful failures now exit 1 (an estimate with too few control points for the initial alignment; a missing --sparse_eval_result); successful runs exit 0.
• evaluate_wrt_mps.py has the same change but was not run, as I don't have MPS data for the demo.
• format.sh with ruff 0.12.7 reports no changes.