Fix use of convdims.

[maleadt]

JuliaGPU/CUDA.jl#873 changed CUDNN.convdims; replicating that change here. This has entered the CUDA 3.1 release, so users will run into this.

That said, NNlibCUDA shouldn't be calling these methods, and instead use the high-level CUDNN API. Currently it has just copied the old implementation that calls into low-level wrappers, while using unexported helper functions like CUDNN.convdims to do so.

[DhairyaLGandhi]

Thank you so much! Should we also copy some tests to catch the instances that the current changes were meant to catch? We can do so in a follow up, but thinking about the correct way forward here.

[maleadt]

I guess NNlibCUDA should call the public cudnnConvolutionForward method, which creates the descriptor based on the other keyword arguments, instead of creating the descriptor manually here. That code is then covered by the CUDA.jl tests.

[DhairyaLGandhi]

Thanks @maleadt and @DrChainsaw

[DhairyaLGandhi]

Makes sense to split the public bits from the descriptor handling.