use the same approach like in fisheye calibration: instead of setting
masked out rows to zero, remove them from the equation system.
This way JtJ does not become singular and we can use the much faster LU
decomposition instead of SVD.

This results in a speedup of the Calibrate unit tests of 3x-10x.

applying patch by rxtsolar: https://github.com/Itseez/opencv/pull/5422 for the master branch (even though it's actually not that important here)

1) fix multichannel support
2) remove useless bad_value, read value from original Mat directly
3) add more tests
4) fix docs for cvCeil and checkRange

WinRT parallels library is causing conflicts in VC++ due to inconsistent usage with C++/CX which is now present in parallels.cpp.  This support should be included in all projects with WinRT specific code except if doing special native C++ tasks.

Fix C++/CX support for all WINRT and bring back native-mode only support which was removed