Fixed crash in SSE implementation of FREAK descriptor when number of pairs is set to something other than 512.

See http://code.opencv.org/issues/3889 for more details

In CUDA 6.0 there was a bug in NPP LUT implementation (invalid results when
src == 255). In CUDA 6.5 the bug was fixed.

Replaced NPP LUT call with own implementation (ported from master branch) 
to be independant from CUDA Toolkit version.

Documentation improvement
Bug in output format for JPG set fixed

The format of the training set can be changed with the `-pngoutput` key.
Output image will be resized to a 640x480 size if greater.

see https://github.com/Itseez/opencv/pull/3137

Conflicts:
	modules/core/src/matmul.cpp