The crash was cased by massive Mat objects leak in NativeCameraView class.

EMBED_CUDA and FORCE_EMBED_OPENCV flags added to cmake macro add_android_project;
INSTALL_CUDA_LIBRARIES option added to OpenCV.mk
opencv_dynamicuda library installation with enabled OPENCV_INSTALL_MODULES flag fixed;
CUDA initialization apportunity added to OpenCVLoader.initDebug();
Tutorial-4-CUDA sample reimplemented with static OpenCV and CUDA initialization.

It's not documented, and it does nothing unless CMake 2.0 compatibility
is enabled (and it isn't):

https://github.com/Kitware/CMake/blob/v2.6.0/Source/cmConfigureFileCommand.cxx

CV_CAP_PROP_SUPPORTED_PREVIEW_SIZES_STRING property is not supported
by all VideoCapture backends. Some backends can return 0.0 or -1.0.

Epoch is gone; now the version component names in C++ and CMake
are consistent (except REVISION corresponds to PATCH).

Conflicts:
	modules/java/generator/src/cpp/VideoCapture.cpp

The paths are defined properly with an escape "\ " but you cannot have an escape and
quotes when piping (otherwise, escapes are understood as 2 characters).
So just remove the quotes.

New version of Java wrappers' generator uses different wrappers for VideoCapture
on Android and desktop to prevent binary compatibility issues.

Linker dependencies to all OpenCV modules are invalid.
We should not include other bindings in this list (like "opencv_python").

Previously, run.py would assume that the opencv_java library is in the
same directory as the tests, which is only true on Windows.

The library path depends on the build configuration, which may not be
known until the actual build (e.g. with the Visual Studio generators),
so it can't be stored in the CMake cache for run.py to read. I didn't
want to hardcode into run.py where the library is on each platform,
either. So that's why I used the current scheme with the properties
file. It also makes running the tests without run.py a little easier.

We don't use it, anyway.

Java inlines static finals if they're defined with a constant expression. In
case of version constants we don't want that to happen, since they obviously
change from version to version. If the user substitutes a different OpenCV
jar without recompiling, we want user code to still have relevant values for
the version constants.

This arranges that by turning constant values into function calls, which no
longer count as a constant expression.

It increases performance on some devices like Nexus4.
Target Android SDK version increased up to 14.

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

Now the features2d class wrappers use composition instead of inheritance.