CUDA implementation wants to convert std::vector<KeyPoint> <-> GpuMat.
There is no direct mapping from KeyPoint (mix of int/float fields)
into cv::Mat element type, so this conversion must be avoided.

Legacy mode is turned back for CUDA builds.

* remove overloads with explicit buffer, now BufferPool is used
* added async versions for all reduce functions

Conflicts:
	modules/core/include/opencv2/core/cuda.hpp
	modules/cudacodec/src/thread.cpp
	modules/cudacodec/src/thread.hpp
	modules/superres/perf/perf_superres.cpp
	modules/superres/src/btv_l1_cuda.cpp
	modules/superres/src/optical_flow.cpp
	modules/videostab/src/global_motion.cpp
	modules/videostab/src/inpainting.cpp
	samples/cpp/stitching_detailed.cpp
	samples/cpp/videostab.cpp
	samples/gpu/stereo_multi.cpp

cudev is used for new device layer

This reverts commit fba72cb6, reversing
changes made to 02131ffb.

Made all STL usages explicit to be able automatically find all usages of
particular class or function.

- fixed printCudaDeviceInfo for new CC
- fixed some compilation errors and warnings
- removed unset command from CMake script
- removed unused std imports

fixed MCVC warnings