Sparse kernels for better performance?

[f-dangel]

The implementation of unfoldNd relies on one-hot convolution. This means the convolution kernels are highly sparse. Hence, the code could run faster when using sparse tensors.

Open questions:

• What is the result of a sparse convolution with a dense input? If it's a dense tensor, that would be good.
• Does using sparse tensors provide a benefit in terms of run time? (related: Add a benchmark #4)

[f-dangel]

Sparse operations are currently not supported in PyTorch. Waiting for progress.