This fixes issue #3412 returning incorrect results for single GpuMat's with a single row, which was a result of an error in #3378.

Unfortunatley because GpuMat's with a single row or column are allocated with cudaMalloc and not cudaMallocPitch and 2D texture objects require pitched memory this PR introduces extra staging memory, when the row/col dimension is 1. This is a quick fix for the issue which only introduces extra overhead for single dimension GpuMat's who's use should not be that common.

A "better" fix would be to rework the TexturePtr class so that

__device__ __forceinline__ R operator ()(index_type y, index_type x) const;

staticly dispatches a call to tex1Dfetch when either x or y is 1, but this will be a more significant change.

Dependant on opencv/opencv#23126

Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

• I agree to contribute to the project under Apache 2 License.
• To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
• The PR is proposed to the proper branch
• There is a reference to the original bug report and related work
• There is accuracy test, performance test and test data in opencv_extra repository, if applicable
  Patch to opencv_extra has the same branch name.
• The feature is well documented and sample code can be built with the project CMake

force_builders=Custom
buildworker:Custom=linux-1
docker_image:Custom=ubuntu-cuda:16.04