replace unv 16-bit multiplication operator with v_mul_wrap since behaviour changed by opencv/opencv#12516

Fixed some inconsistencies in interfaces of different white balance algorithms and improved
the documentation a bit. Also added an option of loading a model during runtime in the
learning-based white balance algorithm. Added a simple tutorial describing how to train a
white balance model.

Implemented a learning-based color balance algorithm, which typically provides
noticeably better quality than grayworld, while being several times slower.
Added a python script for training this algorithm, as well as the benchmarking script
for objective quality evaluation.

Also made some modifications to the grayworld algorithm implementation, including the
16 bit image support and overall acceleration by about 1.7 times (for 8 bit images).