Implement (bilinear) upsampling

[maxfreu]

Hi! For all the DL libraries out there, as well as the image processing packages it would make sense to have GPU-powered upsampling (nearest, bilinear, cubic), including their derivatives. Being able to do this in any number of dimensions would be nice of course, but I think 2D would do it to get going finally. From the Flux side the issue is tackled in this PR. I ported the Caffee 2 kernels for bilinear upsampling from pytorch in this gist. Please have a look and decide if you want to have it here in some form and tell me which changes to make and where to put it eventually. I could then set up a PR, including more rigorous tests. Note that there is an older attempt here.

[maleadt]

Looks good. atomic_add! and pointer operations can be avoided by using @atomic. And I assume it wouldn't be just the kernel that's added to CUDA.jl, but more specifically an implementation of an NNlib interface. If that's the case, CUDA.jl is a good place to implement that interface.