Winograd/Cook/Toom Convolution

[smldis]

If you don't mind I am opening an issue just to track the possibility of having an implementation of the winograd convolution method that is faster than the traditional im2col + gemm method.
See https://arxiv.org/abs/1509.09308.

Also from one of the authors of that paper:
Maratyszcza/NNPACK#8 (comment)