stereosgbm - doing XSOBEL and intensity simultaneously does not make sense (Feature #358)

This is my first ticket for opencv. I hope it is reasonable.

Recently, an XSOBEL filter was added to cvstereosgbm.cpp.

In the literature this filter has been mainly used to remove intensity bias between the image stereo pairs.

If you look at how XSOBEL is used in opencv, especially the line

for( c = 0; c < cn*2; c++, prow1 += width, prow2 += width )

You can see (and debug) that the BT filtered intensity difference costs are added onto the cost cube ALS WELL AS the XSOBEL (and BT) filtered difference costs. This means, you do not get rid of the intensity bias but have just a little "vertical edge enhancement".

I assume it only makes sense to do either pure or XSOBEL filtered intensity. This would either require an additional bool variable or - in the sense of Donald Knuth - to leave XSOBEL out of the SGM routine and build a separate XSOBEL filter in opencv. I could also imagine more filters as the stereo comunity is testing a whole variety of preprocessing filters at the moment. (I would like a 7x9 census, which is the current favourite in the community - which unfortunately requires to modify the pixel difference operation within cvstereosgbm as well.)