On a 32-bit compiler the calculation may result in data (size_t) overflow when running some paralleled algorithms (which can safely run on a 64-bit compiler).
This bug is found when running OpenCV's Retina tutorial on 32bit VS2010.

Additional CMAKE flag WITH_WINRT added.

That way, core/internal.hpp doesn't have to depend on cvconfig.h, which we
don't ship.

This fixes build without TBB and CUDA. Issue #2445

* Refactor auto-detection of parallel frameworks
* Implement cv:getNumThreads, cv::setNumThreads and cv::getThreadNum for all supported frameworks
* From now cv::setNumThreads(0) can be used to turn off parallelisation

added recommended number of stripes to parallel_for_, modified some of the functions to use larger stripes (for better performance)

added cross-platform Mutex implementation; enable platform-native (GDC/Concurrency) parallel_for_ implementation when TBB is not installed.