[utils] Expand grayscale images in rotate_image - #2195

Thanks for the fix! (H, W) images now work with expand=True, including with preserve_origin_shape and at 90°, and match channel 0 of the RGB result. RGB images are unchanged. One thing left before merging:

(H, W, 1) images lose their channel axis

cv2.warpAffine (and cv2.resize) drop a singleton channel axis, so an (H, W, 1) input now comes back as (H', W'). Before this PR it crashed in the aspect-ratio pad. Now that _pad_width uses rot_img.ndim it no longer crashes, but a caller passing an (H, W, 1) mask silently gets a different number of dimensions back. The path without expand already did this, so it's best fixed for all modes at once. Restoring the axis just before returning does that:

            rot_img = cv2.resize(rot_img, image.shape[1::-1], interpolation=cv2.INTER_LINEAR)

    if rot_img.ndim < image.ndim:  # cv2 drops a singleton channel axis
        rot_img = rot_img[..., None]

    return rot_img

With this, img[..., :1] returns (32, 64, 1), (60, 120, 1), (32, 64, 1) and (64, 128, 1) for the no-expand, expand, expand + preserve_origin_shape and 90° cases, and matches channel 0 of the RGB result each time.

Could you also:

• add an (H, W, 1) case to test_rotate_image that checks the trailing axis is kept;
• use a random image for the new grayscale assertions, e.g. np.random.rand(32, 64, 3).astype(np.float32). With an all-ones img, np.array_equal(gray, color[..., 0]) would also pass if values were mixed across channels or the image were flipped;
• mention the accepted shapes in the docstring: image: array of shape (H, W) or (H, W, C).