feat: page-aware mAP scoring (thread page_aware through the pipeline)

[Divya-Bhargavi]

Summary

Threads page-awareness through the mAP scoring pipeline, building on the
page_aware BBoxIoUComparator from this base branch (sr/bbox-page-aware).

That branch made the comparator page-aware, but the end-to-end mAP path still
constructed a plain BBoxIoUComparator(), so a prediction with correct
coordinates on the wrong page would still score as a hit in mean_ap. This
PR closes that gap.

Changes

• MAPCalculator(page_aware=...) — passes the flag to its
  BBoxIoUComparator.
• BBoxMAPAccumulator(page_aware=...) — forwards to MAPCalculator for the
  bulk path.
• compare_with(add_bbox_metrics=True, bbox_page_aware=...) (and the
  StructuredModel wrapper) — single-document path.

Default stays page_aware=False, so existing behavior is unchanged. No
data-model change was needed: page-suffixed boxes
([[x1, y1], [x2, y2], page] / [x1, y1, x2, y2, page]) already flow from
_bbox extras straight to the comparator.

Behavior

With page_aware=True, a field is a localization hit only when the page matches
and IoU ≥ threshold:

GT page	Pred page	Coords	page-aware mAP
1	1	identical	1.0
1	2	identical	0.0 (wrong page)
1	(none)	identical	0.0 (page-less is a miss)

Page-unaware (default) ignores the page and scores the geometry only.

Usage

# Bulk
evaluator = BulkStructuredModelEvaluator(
    accumulators=[BBoxMAPAccumulator(page_aware=True)],
)

# Single document
result = gt.compare_with(pred, add_bbox_metrics=True, bbox_page_aware=True)

with page-suffixed rich values, e.g.
{"_value": "Acme", "_bbox": [[10, 20], [200, 50], 1]}.

Testing

• 11 new tests across the calculator, compare_with, and bulk accumulator
  paths (same-page hit, wrong-page miss, page-less miss, default-unaware).
• Full suite: 1452 passed, 2 skipped.
• ruff check and ruff format --check clean on changed files.

Stacked on sr/bbox-page-aware; once that merges, this can retarget to dev.

By submitting this pull request, I confirm that you can use, modify, copy, and
redistribute this contribution, under the terms of your choice.