Convolution calculation now uses 2x transposing horizontal pass
instead of horizontal and vertical which simplifies the code by
reducing code duplication.

Instead of more convolution methods, only one generic method is
used and kernels are now a parameter to the method. For this
Kernel class was introduced with responsibility to calculate
kernel walues and hold kernel specific information.

Besides Lanzcos resampling also Box, Bilinear and Bicubic resampling
was introduced. Box is the simplest and fastest but with still
good quality.