Fix wrong COCO bbox area in vector object-detection export

[jaideeppyne]

Problem

In sa_vector_to_coco_object_detection (SA → COCO vector export), the annotation area is computed as:

area = int((points["x2"] - points["x1"]) * points["y2"] - points["y1"])

Operator precedence parses this as ((x2 - x1) * y2) - y1, not the COCO bounding-box area (x2 - x1) * (y2 - y1) (width × height). The missing parentheses around (y2 - y1) make the error term exactly y1 * (width - 1), so the exported area is wrong for every bbox that isn't flush against the top edge (y1 == 0).

On this repo's own export golden box {x1: 437.16, y1: 341.5, x2: 465.23, y2: 357.09}:

• current code → area = 9682
• correct w * h → 437 (which is exactly what the shipped golden TestVectorAnnotationImage.json lists)

The sibling paths already do it right — instance-segmentation uses int(_area(mask)) and keypoint uses bbox[2] * bbox[3] — so the intent is unambiguous.

Fix

-        area = int((points["x2"] - points["x1"]) * points["y2"] - points["y1"])
+        area = int(bbox[2] * bbox[3])

bbox is already (x, y, width, height), so this reuses the computed width/height and mirrors the keypoint path.

Tests

Adds tests/unit/test_coco_object_detection_area.py, asserting area == int(width * height) (and == 437 for the golden box). It fails on the current code (assert 9682 == 437) and passes with the fix.

The existing export tests never caught this because they only assert output filename equality (filecmp.dircmp left_only/right_only) and never compare file contents — so the wrong area values were never checked.