feat: add itkwasm-downsample-cucim

[thewtex]

This is based on the itk_cucim package work. _discrete_gaussian.py by Greg Lee is copied directly.

We are currently getting sporatic differences in the test output of 1.

[thewtex]

@grlee77 @gigony please take a look 🙏 . I see affine_transform in cucim: https://github.com/rapidsai/cucim/blob/d4ec01eb53afceaaeec36856e1dcbf2c282659d0/python/cucim/src/cucim/skimage/_vendored/_ndimage_interpolation.py#L344-L356 but how to import it? Which is better to use -- that or the one in cupy?

The motivation is to provide NVIDIA acceleration when available for OME-Zarr generation per thewtex/ngff-zarr#68

[thewtex]

@tbirdso @agirault @jjomier per our discussion at the hackathon, we can use GPU representations on the image.data pixel buffer. In this case, it is a cupy array. This will reside on the GPU until needed in the CPU memory. In the test here, that is happening in write_image.

[thewtex]

Failing CI is unrelated.