Add __cuda_array_interface__ to itk::CudaImage

[thewtex]

A __cuda_array_interface__ method could prevent CPU/GPU memory transfer when itk.CudaImage is used with cupy, pytorch, etc.

@gigony you shared a nice reference of how this is implemented in cuCIM, but I am having a hard time finding it in a quick grep of the cucim repository. Any pointers?

Edit: Found it :-):

https://github.com/rapidsai/cucim/blob/4061b7e00c5b10f18f04f3c356b32a42c081de4f/python/pybind11/cucim_py.cpp#L628-L631

CC @SimonRit @LucasGandel

[gigony]

Hi @thewtex

For supporting __cuda_array_interface__, you can reference this code implemented in cuCIM.

• https://github.com/rapidsai/cucim/blob/main/python/pybind11/cucim_py.cpp#L627
• https://github.com/rapidsai/cucim/blob/main/python/pybind11/cucim_py.cpp#L534

[jakirkham]

Also here's some code from CuPy for consuming/producing __cuda_array_interface__ objects:

• Consuming: https://github.com/cupy/cupy/blob/10d555d64d5676007f06d732f83584dc916d406a/cupy/_core/core.pyx#L2652
• Producing: https://github.com/cupy/cupy/blob/10d555d64d5676007f06d732f83584dc916d406a/cupy/_core/core.pyx#L192

[LucasGandel]

Soon addressed by #42.
Thank you all for your valuable insight ❤️ 🚀 💯