Share box conversion kernels between ops and transforms v2

[YeonwooSung]

ops.box_convert and F.convert_bounding_box_format both converted boxes, but the arithmetic lived in two places. This keeps the kernels in torchvision/ops/_box_convert.py and has transforms v2 call them.

What is now shared

• Axis-aligned and rotated conversion math lives in ops/_box_convert.py (_box_* helpers plus _half_wh).
• ops.box_convert uses those helpers directly, including new first-class xywh ↔ cxcywh kernels instead of a two-step trip through xyxy.
• v2 convert_bounding_box_format calls the same kernels for the arithmetic.
• xyxy → cxcywh (and xywh → cxcywh) use overflow-safe origin + size/2 instead of (x1 + x2) / 2 or (x1 * 2 + width) / 2. This does not fight Avoid overflow in box center conversion #9641; it uses the same safer form.

Backward compatibility

• ops.box_convert is still out-of-place, still JIT-scriptable, and still true-divides (integer inputs may promote to float).
• v2 still supports inplace=True and still keeps integer dtype via floor-div on the xyxy/xywh → cxcywh paths.
• Integer CXCYWH → XYWH still recomputes w/h after truncation so it matches the two-step v2 path (PR Fix CXCYWH to XYXY conversion for integer bounding boxes #9322 / issue Incorrect BoundingBoxes format conversion CXCYWH -> XYXY #8887).
• ops does not import transforms.

Tests

• ops.box_convert and F.convert_bounding_box_format are checked to agree on float32/float64 for xyxy/xywh/cxcywh pairs.

Fixes #9387

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9651

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.