clip_boxes_to_image can drop valid boxes for large coordinates

[Rajioba1]

Describe the bug
spatial_crop_boxes converts roi_start and roi_end to torch.int16 before clamping boxes. For coordinates >= 32768, the ROI bounds overflow/wrap, so boxes that are fully inside the requested crop can be clamped to zero width and dropped when remove_empty=True.

This is also reachable through the public clip_boxes_to_image API for large images.

To Reproduce

import torch
from monai.data.box_utils import clip_boxes_to_image, spatial_crop_boxes

boxes = torch.tensor([[41000.0, 5000.0, 45000.0, 15000.0]], dtype=torch.float32)

cropped, keep = spatial_crop_boxes(
    boxes,
    roi_start=[40000, 0],
    roi_end=[50000, 20000],
    remove_empty=True,
)
print(keep.tolist(), cropped.tolist())

clipped, keep = clip_boxes_to_image(boxes, spatial_size=[50000, 50000], remove_empty=True)
print(keep.tolist(), clipped.tolist())

Current output:

[False] []
[False] []

Expected behavior
The box is inside the crop/image and should be kept. For the crop example, the expected cropped box is:

[True] [[1000.0, 5000.0, 5000.0, 15000.0]]

For the clip_boxes_to_image example, the expected clipped box is unchanged and kept.

Environment

MONAI version: 1.6.0rc1+48.g8690ae74
Numpy version: 1.26.4
Pytorch version: 2.10.0+cpu
MONAI rev id: 8690ae74a8a489d31fe1f9ac8ef0bff63165383e
System: Windows-11-10.0.26200-SP0
Python version: 3.12.1

Additional context
The overflow appears to come from the .to(torch.int16) casts on the converted ROI bounds in monai/data/box_utils.py. Whole-slide pathology and detection workflows can use image coordinates larger than 32767, so this can silently remove valid detections instead of raising an error.