[yuriyzubov]

Summary

Builds on PR #210. Reworks instance scoring to emit per-instance Hausdorff
sum/count (instead of preaggregated values), computes per-class combined scores
from pooled Hausdorff, and splits results output into a public aggregate file
and a full _extended sibling.

Changes

Scoring

• score_instance emits per-instance Hausdorff sum/count
• _compute_hausdorff_scores returns a per-instance array
• _create_pathological_scores takes only status
• empty_label_score penalizes non-submission (worst-case boundary)

Distance

• vectorize normalize_distance (scalar or array)
• Clamp ROI pad to volume extent (fixes inf-cast warning)

Aggregation / output

• Per-class combined score from pooled Hausdorff; plain-mean overalls
• Write public aggregate-only results to result_file
• Write full results to an _extended sibling file

Types

• InstanceScoreDict uses hausdorff_norm_sum/n_hausdorff

Tests

• Cover public-scores filtering and combined-score pooling across crops
• Instance/Hausdorff scoring tests updated to tp/fn/n_hausdorff

Notes

Stacked on PR #210