Perf: skip redundant full-image mask on StdShiftIntensity nonzero=False path (#8975)

Fixes #8972 .

### Description

On the default `nonzero=False` path `StdShiftIntensity._stdshift` built
an all-True boolean mask the size of the image and shifted through it
(`img[slices] = img[slices] + offset`), forcing a full advanced-index
gather and scatter plus the mask allocation and `.any()` scan even
though every voxel is selected. That is equivalent to shifting the image
directly, so the `nonzero=False` branch now does `img + factor *
std(img)`. Output is bit-for-bit identical and the `nonzero=True` path
is untouched. `RandStdShiftIntensity` benefits too, since it also
defaults to `nonzero=False`. Measured across 2D/3D, single and multi
channel, float32 and float64 on both backends (best-of-3, CPU); output
verified equal in every configuration:

| shape | dtype | backend | current (ms) | proposed (ms) | speedup |
|---|---|---|---|---|---|
| 1x256x256 | f32 | numpy | 0.123 | 0.048 | 2.58x |
| 1x256x256 | f32 | torch | 0.563 | 0.039 | 14.47x |
| 1x1024x1024 | f32 | numpy | 3.016 | 1.294 | 2.33x |
| 1x1024x1024 | f32 | torch | 7.040 | 0.274 | 25.66x |
| 1x1024x1024 | f64 | torch | 9.644 | 0.601 | 16.04x |
| 1x64x64x64 | f32 | numpy | 0.637 | 0.251 | 2.54x |
| 1x64x64x64 | f64 | torch | 2.646 | 0.095 | 27.84x |
| 1x128x128x128 | f32 | numpy | 7.019 | 2.914 | 2.41x |
| 1x128x128x128 | f32 | torch | 41.618 | 0.630 | 66.09x |
| 1x128x128x128 | f64 | torch | 44.534 | 1.825 | 24.40x |
| 4x96x96x96 | f32 | torch | 68.152 | 1.357 | 50.23x |
| 4x160x160x160 | f32 | numpy | 128.429 | 82.662 | 1.55x |
| 4x160x160x160 | f32 | torch | 342.641 | 16.727 | 20.48x |
| 4x160x160x160 | f64 | torch | 354.009 | 41.652 | 8.50x |

numpy ranges 1.55x to 2.65x and torch 8.5x to 66x across the full sweep;
the largest gains are on torch, where all-True boolean-mask indexing is
especially costly.

### Types of changes
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).

---------

Signed-off-by: Soumya Snigdha Kundu <soumya_snigdha.kundu@kcl.ac.uk>
Co-authored-by: Eric Kerfoot <17726042+ericspod@users.noreply.github.com>

Author: aymuos15

monai/transforms/intensity/array.py

@@ -354,17 +354,14 @@ def __init__(
         self.dtype = dtype
 
     def _stdshift(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
-        ones: Callable
         std: Callable
         if isinstance(img, torch.Tensor):
-            ones = torch.ones
             std = partial(torch.std, unbiased=False)
         else:
-            ones = np.ones
             std = np.std
 
-        slices = (img != 0) if self.nonzero else ones(img.shape, dtype=bool)
-        if slices.any():
+        slices = (img != 0) if self.nonzero else ()
+        if not self.nonzero or (isinstance(slices, (np.ndarray, torch.Tensor)) and slices.any()):
             offset = self.factor * std(img[slices])
             img[slices] = img[slices] + offset
         return img