Workaround for performance regression introduced by FFTW

[galenlynch]

Convolutions in DSP currently rely on FFTW.jl, and a recent change in FFTW.jl
(JuliaMath/FFTW.jl#105) has introduced a large performance regression in conv
whenever Julia is started with more than one thread. Since v1 of FFTW.jl, it uses multi-threaded
FFTW transformations by default whenever Julia has more than one thread. This
new default causes small FFT problems to run much more slowly and use much more
memory. Since the overlap-save method of conv in DSP breaks a convolution
into many small convolutions, and therefore performs a large number of small FFTW
transformations, this change can cause convolutions to be slower by two orders
of magnitude, and similarly use two orders of magnitude more memory. While
FFTW.jl does not provide an explicit way to set the number of threads used by a
FFTW plan without changing a global variable, generating the plans with the
planning flag set to FFTW.PATIENT (instead of the default MEASURE) allows
the planner to consider changing the number of threads. Adding this flag to the
plans generated by the overlap-save convolution method seems to rescue the
performance regression on multi-threaded instances of Julia.

Fixes #339
Also see JuliaMath/FFTW.jl#121

[martinholters]

This seems to take significant time when running the tests---enough for travis to cancel the jobs after 10 minutes without output.

[martinholters]

The planning time seems to become excessive in the higher-dimensional case:

julia> x = zeros(128, 128, 128);

julia> out = zeros(255, 255, 255);

# master
julia> @time DSP.unsafe_conv_kern_os!(out, x, x, size(x), size(x), size(out), (256, 256, 256));
  9.859362 seconds (13.47 M allocations: 1.152 GiB, 3.15% gc time)

julia> @time DSP.unsafe_conv_kern_os!(out, x, x, size(x), size(x), size(out), (256, 256, 256));
  5.168003 seconds (2.27 k allocations: 515.127 MiB, 2.97% gc time)

# this PR
julia> @time DSP.unsafe_conv_kern_os!(out, x, x, size(x), size(x), size(out), (256, 256, 256));
751.674809 seconds (13.06 M allocations: 1.381 GiB, 0.04% gc time)

julia> @time DSP.unsafe_conv_kern_os!(out, x, x, size(x), size(x), size(out), (256, 256, 256));
  3.693583 seconds (2.25 k allocations: 772.126 MiB, 4.53% gc time)

(Note: this is without multiple julia threads.)

This looks prohibitive, I'm afraid, but I don't have an alternative idea.

[galenlynch]

Ah, bummer :(

I guess we'll have to wait to see if JuliaMath/FFTW.jl#150 gets merged...