Add product DFTs to gr_dft; make acb_dft a thin wrapper

[fredrik-johansson]

We change acb_dft to a thin wrapper around gr_dft for the main functions, removing obsolete algorithm functions. This makes acb_dft significantly faster (and probably more precise). The only significant functionality that was missing, acb_dft_prod, is also ported to gr_dft, and with it acb_dirichlet_dft.

@edgarcosta Would be great if you could test accuracy/speed of the new DFT code for the applications where you've been using it.

Also @Joel-Dahne maybe you are able to test this.

As a sanity check, this immediately speeds up vector evaluation of Dirichlet L-functions.

Old:

$ build/examples/lcentral --quiet --prec 32 1000000 1000000
cpu/wall(s): 1.939 1.939
virt/peak/rss/peak: 38.8 MB /  149 MB / 27.4 MB /  137 MB

New:

$ build/examples/lcentral --quiet --prec 32 1000000 1000000
cpu/wall(s): 0.692 0.693
virt/peak/rss/peak: 20.5 MB /  131 MB / 8.59 MB /  118 MB

Enclosures are also tighter. Old:

...
1000000,810547: [0.73711 +/- 7.86e-6] + [-0.68267 +/- 6.26e-6]*I
1000000,199219: [4.79556 +/- 4.50e-6] + [3.88048 +/- 5.39e-6]*I
1000000,976563: [1.49068 +/- 4.92e-6] + [1.27756 +/- 7.06e-6]*I

New:

...
1000000,810547: [0.737105 +/- 6.65e-7] + [-0.682667 +/- 7.97e-7]*I
1000000,199219: [4.79556 +/- 1.95e-6] + [3.880482 +/- 8.42e-7]*I
1000000,976563: [1.490678 +/- 8.34e-7] + [1.277564 +/- 6.90e-7]*I

[Joel-Dahne]

This looks great! There were no problems wrapping it in Julia (though it needs some polishing to make it actually use friendly) and the performance seems to be very good! I made a quick comparison to GenericaFFT.jl and at 53 bits of precision it is about 50 times slower than Float64, which or more or less in line what I would expect. Compared to BigFloat at 256 bits of precision it is about 20 times faster, though GenericFFT is suffering a lot from not using mutable arithmetic with BigFloat. All tests were for the length 32768.