OpenCV (with CUDA enabled) should only use Cuda Runtime API functions (Feature #3981)

I have compiled OpenCV 2.4.9 with CUDA enabled (Cuda Toolkit 5.0) with Visual Studio 2008 (64bit) in Windows. When I was checking with the Windows Dependency Walker the dependencies of the 'opencv_core' DLL, I noticed that in OpenCV (at least in the 'opencv_core' library) unfortunately one function (cuDeviceGetAttribute) from the CUDA driver API is used. Due to this, the 'opencv_core' DLLT has a dependency against 'nvcuda.dll'. This is a BIG problem if you want to deploy this OpenCV variant on machines with either no CUDA GPU or on virtual machines. Because on those systems, 'nvcuda.dll' will not be found and the application will NOT start, even if NO GPU-accelerated OpenCV function is used within the application !

See also https://stackoverflow.com/questions/26509111/opencv-with-cuda-deployable-on-systems-without-cuda-capable-gpu

My proposal would be to use the equivalent function 'cudaGetDeviceProperties' from the CUDA runtime API instead of the fn 'cuDeviceGetAttribute' . By this, appliciations using the OpenCV variant (with CUDA enabled) will run also on machines with NO CUDA GPU installed (if no GPU-accelerated function is called). Note the alternative - the usage of two different OpenCV libraries (with CUDA, without CUDA) is software-technically NOT desired in our department.