Add OpenCL  support to linearPolar & logPolar.
The OpenCL code use float instead of double, so that it does not require
cl_khr_fp64 extension, with slight precision lost.

Add explicit conversion

Add explicit conversion from double to float to eliminate warning during
compilation.

templated std::vector<T> constructor was being called instead of const std::vector<cv::cuda::GpuMat>.

Rewrite linearPolar & logPolar so that they do not depend on the
deprecated API CvMat. Issue 6377 is resolved in this way because the two
routines do not convert UMat to CvMat anymore.

They were issued for every frame retrieved - even in a release build.

"Eat" unexpected backend exceptions

Added 2 inline functions in persistence.hpp
So that the '>>' operator works correctly for std::Vector<KeyPoint> and
std::Vector<DMatch>