See pull request #1241 for why.

do not calculate sum of error in every round of iteration;
instead the error will be summed every 2nd times or more, 
if the previous sum of error is too far away from threshold.

* use explicit extrapolation only for sub-matrixes
* use built-in interpolation for INTER_NEAREST mode

Changed the brief description of function Mat::colRange from "Create a matrix header for the specified row span." to "Create a matrix header for the specified column span."

On a 32-bit compiler the calculation may result in data (size_t) overflow when running some paralleled algorithms (which can safely run on a 64-bit compiler).
This bug is found when running OpenCV's Retina tutorial on 32bit VS2010.