When using "valid" convolutions with subsampling, the gradient
implementation was wrong (if dx or dy > 2) or slow.
In that case, we now use the implementation in conv3D instead.