cudev - Rework some code

[nglee]

• Use shfl_down, instead of __shfl_down, on scan.hpp

modules/cudev/include/opencv2/cudev/warp/scan.hpp
modules/cudev/include/opencv2/cudev/warp/shuffle.hpp

This fix is applied to make the code conform to the structure of cudev module. __shfl_down should be directly called at shuffle.hpp and other files in cudev module should call shfl_down, which is a wrapper for __shfl_down.

• Remove race conditions

modules/cudev/include/opencv2/cudev/block/scan.hpp
modules/cudev/include/opencv2/cudev/grid/detail/integral.hpp

This fix is applied to pass the following cuda-memcheck --tool racecheck tests:

cuda-memcheck --tool racecheck opencv_test_cudev
cuda-memcheck --tool racecheck opencv_test_cudaarithm --gtest_filter=*Integral*

force_builders=Custom
buildworker:Custom=linux-1,linux-2,linux-4
docker_image:Custom=ubuntu-cuda:16.04

[alalek]

Thank you!