Disparity Refinement

I'd like to add a refinement step of the disparities after the [winner_takes_all](
https://github.com/fixstars/libSGM/blob/master/src/stereo_sgm.cpp#L185) step. Therefore, I'd need to know at which line of code the acutual matching costs per pixel are obtained.

The winner_takes_all kernel receives the [smoothed costs](https://github.com/fixstars/libSGM/blob/master/src/winner_takes_all.cu#L41) (d_scost) and somewhere at the end of the kernel the disparity is calculated for [left](https://github.com/fixstars/libSGM/blob/master/src/winner_takes_all.cu#L113) (leftDisp) and right (rightDisp). I'm somewhat confused where the actual matching costs are obtained. I'm guessing it could be [lhv](https://github.com/fixstars/libSGM/blob/master/src/winner_takes_all.cu#L112) or [minCostL2](https://github.com/fixstars/libSGM/blob/master/src/winner_takes_all.cu#L112).

Read src/stereo_sgm.cpp around line 185 and src/winner_takes_all.cu around lines 41 and 112-113. Trace d_scost, lhv, minCostL2, leftDisp, and rightDisp to determine where per-pixel matching costs are available; done means clarifying the correct cost values and the location and scope of the disparity-refinement step.

Written by the indexing model from the issue text.