This provides a great speedup, especially the relationship between kernel
length and processing size is now logarithmic instead of linear. Below a
kernel size of 32 it's a bit slower, afterwards it's much faster:

17     0.788000 -> 0.950000
33     1.208000 -> 1.146000
65     2.166000 -> 1.146000
...
4097 107.444000 -> 1.508000

For sizes smaller 32 the normal time-domain convolution is chosen,
for larger sizes the FFT convolution is automatically used.

Fixes bug #594381.