Add strides, implement CUDA Array Interface

[maleadt]

We should probably implement the CUDA array interface for interoperability across deep learning frameworks:

• https://developer.nvidia.com/blog/machine-learning-frameworks-interoperability-part-1-memory-layouts-and-memory-pools/
• https://numba.pydata.org/numba-doc/dev/cuda/cuda_array_interface.html

As part of that, we'll need to a dd a strides field to CuArray. This should make it possible to get rid of some more SubArray uses, as well as do some more advanced optimization like under-Peter/OMEinsum.jl#133.

[pabloferz]

I would advertise DLPack.jl as a solution for this. It already gives you CuArrays and can interoperate with most python libraries that have an array implementation.

[maleadt]

Awesome, I didn't realize that already supported CUDA.jl. Numba is still using the array interface though, numba/numba#4719. Since I don't plan on working on this, I'll go ahead and close the issue. If anybody cares, we can re-open and discuss how to move forward here.