Nufft scaling on CPU for large thread counts

[mreineck]

As mentioned in the last meeting, I had the chance of doing a few NUFFT scaling tests on a very large node (AMD EPYC 9965, 384 physical cores in total, AVX512 support).
I compiled the latest versions of Finufft and ducc from source and ran type 1/2 double precision transforms on a 256**3 grid with varying nonuniform point densities (1 and 10) and thread counts (32, 64, 128).
Here are a few results (Finufft was compiled with ducc FFTs):

Density 1:

Density 10:

I also tried Finufft with the FFTW backend, but it seems that FFTW goes a little berserk when confronted with many threads:

Some lessons learned:

• FFTW has problems with large thread counts (at least when FFTW_ESTIMATE is used).
• Scaling with number of threads is far from optimal in both libraries (this is also a reason why I didn''t try to scale to all 384 cores). Not sure if this will get significantly better for larger problems ...
• The scaling of Finufft's planning stage could be improved.
• Performance at epsilon=1e-6 is not much different from the A100 GPU results reported by @jipolanco, which I'm quite happy about :-)