From commit dd74a851, to be exact. Now cap_ffmpeg.cpp should actually
build if HAVE_FFMPEG is true.

Also modified some gpu sources in a similar manner.

It supports NORM_L1 distance types now and can
use user provided indices.
Also fixed a bug of kmeans where distance pointers should be float instead
 of double.

NORM_L2 changed to NORM_L2SQR, Accuracy and Perf tests are added

added ROI support in accuracy test of distanceToCenters

Need to go back at some point and fix this for real.

(cherry picked from commit dfe07df8)

Newer FFmpeg prohibits 4096x4096 MPEG-2, presumably because it violates
the standard.

http://git.videolan.org/gitweb.cgi/ffmpeg.git/?p=ffmpeg.git;a=commit;h=7fb87bc5f24b1be13269109506c05e4c54695b5e

(cherry picked from commit 424a7b0a)

Conflicts:
	modules/highgui/test/test_ffmpeg.cpp

(cherry picked from commit 3f3ae333)

While the FFmpeg video IO backend gets compiled even without it,
the tests (for both FFmpeg and video IO) don't. This should fix
that.

The reasons for that are twofold:

1) LINK_PRIVATE is only available since CMake 2.8.7.
2) The way it was used generated a warning because of CMake policy CMP0023:

   http://www.cmake.org/cmake/help/v2.8.12/cmake.html#policy:CMP0023

Using LINK_INTERFACE_LIBRARIES actually causes another warning - this time
because of CMake policy CMP0022:

   http://www.cmake.org/cmake/help/v2.8.12/cmake.html#policy:CMP0022

I set the policy to OLD, because NEW means subtle changes when compiling
with CMake 2.8.12, and I don't want to research that this close to release.
:-)

I also removed the setting of CMP0003, because it's set by
cmake_minimal_version anyway.

It was actually acting as an additional argument, breaking the error message.