Faster cropping and scaling of images

## ❓Question
I am trying to predict some features of a very small object in a very large image. To make it manageable I therefore go for a 2-step approach that takes a full resolution crop of the object, simplified like this:

```python
import torch
from torchvision.models import mobilenet_v3_large

model_1 = mobilenet_v3_large(num_classes=2)
model_2 = mobilenet_v3_large()
image = torch.randn((1, 3, 3840, 2160))

small_image = image[:, :, 16:-16:17, 9:-9:17] # downscale
center_yx = model_1(small_image).int() # find object in image
crop = image[:, :, center_yx[0]-112: center_yx[0]+112, center_yx[1]-112: center_yx[1]+112] # crop
features = model_2(crop) # predict features
```

This works well except it is very slow when converted to CoreML/ML program. When testing the model performance through Xcode it says that about 75% of the time is spent on image cropping and downscaling (`ios18.slice_by_index`).

Is `slice_by_index` expected to be this slow? Are there any faster alternatives to scaling and cropping images? Or any recommend way of tackling the problem?

Start with the Python reproduction using mobilenet_v3_large and the converted Core ML model, then inspect the Xcode performance report for ios18.slice_by_index. The issue needs an explanation of whether cropping and downscaling cause the reported overhead and, if possible, a supported faster approach; no specific test or acceptance criterion is provided.

Written by the indexing model from the issue text.