Fix a bug in the er tree clean-up process (see discussion in https://github.com/Itseez/opencv/pull/1520). It's not necessary to walk the er tree in order to clean up the memory since it's possible to cleanup the regions in the stack just by iterating over the er_stack vector instead of walking the tree structure. This turns out to be safer because the tree structure of those regions may have been modified previously in er_save function.

Updated documentation and sample to use the new function API

Added erGrouping function: Find groups of Extremal Regions that are organized as text blocks. Updated sample/cpp to use the complete text detection pipeline

first parameter of createERFilterNM1/createERFilterNM2 is now mandatory. changed the sample program to use the new prototypes

renamed getDefaultClassifierNM{1,2}() to loadClassifierNM{1,2}() and added "const std::string& filename" parameter (mandatory).

functions getDefaultClassifierNM1() and getDefaultClassifierNM2() allow to implicitly load the default classifiers when creating a ERFilter object

Using cv::magnitude instead of custom loop and Mat<float> instead of Mat<double> to compute gradient magnitude channel.

added computeNMChannels helper function to compute the diferent channels to be processed independently in the N&M algorithm.

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

Neumann L., Matas J.: Real-Time Scene Text Localization and Recognition, CVPR 2012.

High-level C++ interface and implementation of algorithm is in the objdetect module.
C++ example, a test image, and the default classifiers in xml files.