perf(RandCropBoxByPosNegLabeld): avoid N-fold mask allocation in randomize (#8993)

### Description

`RandCropBoxByPosNegLabeld.randomize` builds a foreground/background
mask when precomputed indices are not available. The previous
implementation called `convert_box_to_mask` for each box (creating an
N-channel stacked mask) then reduced with `np.amax` to a single channel,
allocating an (N_boxes, H, W, D) int16 intermediate that grows linearly
with the number of boxes.

The fix paints directly into a single (1, H, W, D) union mask by
iterating over each extended box and setting `mask[0, box_slices] = 1`.

### Changes

**`monai/apps/detection/transforms/dictionary.py`**

1. Remove the now-unused import of `convert_box_to_mask` from
`monai.apps.detection.transforms.box_ops`.

2. Replace the `convert_box_to_mask` + `np.amax(...)[0:1]` block in
`randomize()` with direct slicing into a pre-allocated zero mask.

### Performance

Benchmarked old (`convert_box_to_mask` + `np.amax`) vs new (direct
slicing) implementations on synthetic 3D volumes with random boxes
(`spatial_size=(64,64,64)`, `num_samples=4`, `pos=1`, `neg=1`,
`whole_box=True`).

**System specs**
- CPU: 12th Gen Intel Core i7-12800H (14 cores / 20 threads, up to 4.8
GHz)
- RAM: 30 GiB
- OS: Ubuntu 22.04.5 LTS, kernel 6.8.0-124-generic
- Python 3.10.12, NumPy 2.2.6, PyTorch 2.13.0+cu130

**`randomize()` only**

| image_size | n_boxes | old (ms) | new (ms) | speedup |
|---|---|---|---|---|
| (96,96,96) | 5 | 2.471 | 1.331 | 1.86x |
| (96,96,96) | 20 | 17.998 | 1.921 | 9.37x |
| (160,160,160) | 5 | 22.931 | 15.241 | 1.50x |
| (160,160,160) | 20 | 78.651 | 6.563 | 11.98x |
| (256,256,128) | 5 | 62.477 | 26.386 | 2.37x |
| (256,256,128) | 20 | 174.831 | 24.427 | 7.16x |

**Full `__call__()`**

| image_size | n_boxes | old (ms) | new (ms) | speedup |
|---|---|---|---|---|
| (96,96,96) | 5 | 6.416 | 4.823 | 1.33x |
| (96,96,96) | 20 | 21.745 | 5.078 | 4.28x |
| (160,160,160) | 5 | 22.820 | 13.778 | 1.66x |
| (160,160,160) | 20 | 85.259 | 14.000 | 6.09x |
| (256,256,128) | 5 | 76.623 | 38.455 | 1.99x |
| (256,256,128) | 20 | 188.849 | 36.955 | 5.11x |

### Types of changes

- [x] Non-breaking change (fix or new feature that would not break
existing functionality).
- [ ] Breaking change (fix or new feature that would cause existing
functionality to change).
- [ ] New tests added to cover the changes.
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [x] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [ ] In-line docstrings updated.
- [ ] Documentation updated, tested `make html` command in the `docs/`
folder.

---------

Signed-off-by: Soumya Snigdha Kundu <soumya_snigdha.kundu@kcl.ac.uk>

Author: aymuos15

monai/apps/detection/transforms/dictionary.py

@@ -37,7 +37,6 @@
     StandardizeEmptyBox,
     ZoomBox,
 )
-from monai.apps.detection.transforms.box_ops import convert_box_to_mask
 from monai.config import KeysCollection, SequenceStr
 from monai.config.type_definitions import DtypeLike, NdarrayOrTensor
 from monai.data.box_utils import COMPUTE_DTYPE, BoxMode, clip_boxes_to_image
@@ -1154,10 +1153,13 @@ def randomize(  # type: ignore
             # As along as the cropped patch contains a box, it is considered as a foreground patch.
             # Positions within extended_boxes are crop centers for foreground patches
             extended_boxes_np = self.generate_fg_center_boxes_np(boxes, image_size)
-            mask_img = convert_box_to_mask(
-                extended_boxes_np, np.ones(extended_boxes_np.shape[0]), image_size, bg_label=0, ellipse_mask=False
-            )
-            mask_img = np.amax(mask_img, axis=0, keepdims=True)[0:1, ...]
+            spatial_dims = len(image_size)
+            mask_img = np.zeros((1,) + tuple(image_size), dtype=np.int16)
+            for b in range(extended_boxes_np.shape[0]):
+                slicing = (0,) + tuple(
+                    slice(extended_boxes_np[b, d], extended_boxes_np[b, d + spatial_dims]) for d in range(spatial_dims)
+                )
+                mask_img[slicing] = 1
             fg_indices_, bg_indices_ = map_binary_to_indices(mask_img, thresh_image, self.image_threshold)
         else:
             fg_indices_ = fg_indices