I have added an overlap-save algorithm to convolution, which is substantially
faster (~2x) for large convolution problems. It is also substantially faster
than other packages like ImageFiltering.jl. For smaller convolutions, it is
about the same or a little bit slower than standard fft convolution. The cutoff
between overlap-save and simple fft convolution is pretty arbitrary at the
moment, so it would be nice to add some benchmarking to allow DSP.jl to choose
between the two algorithms depending on the size of the input. It would also be
nice to add spatial/time domain filtering for small convolutions, as it would
almost certainly be faster than fft convolutions. This is relevant to JuliaDSP#224.

performs overlap-save on the first argument if it is large
enough to warrant it. Performs fft convolution of all others.

public conv needs 2 or more arguments
redundant union removed
permsort with rev=true
comprehension replaced with elementwise operation

(fix minor issue)

also fix issue that test revealed

copy only out array for complex

and test that it runs an ordinary convolution