Fix application of Gaussian Blur with FFTW (about 30% faster)

The patch fixes a slowdown we introduced when implemented
two-stage gaussian convolution application. This two-stage
approach is good for spacial convolution algorithm, but it
is no good for FFTW one. Because every kernel (even one-
dimensional) should be first split into 2D-wave-plane, which is
extremely slow.

CCBUG:407062