initial upsampfact tweak

[DiamonDinoia]

I implemented an heuristic to better choose upsampling factor between 1.25 and 2.00.

I gather the data using devel/wisdom.py then I analyze with devel/analyse_upsamp.py

With that I developed the cases that go in the c++ code.

[mreineck]

This goes far above and beyond what I originally had in mind, thanks!
One thing to keep in mind is that, for a given problem and accuracy, using oversampling factor 1.25 instead of 2.0 may decrease the overall accuracy of the result; this problem especially occurs with

• accuracies close to the machine epsilon (so it's mostly a concern for single precision), and
• increasing dimensionality

So whenever the benchmarks suggest to switch to 1.25 where the current code uses 2, some extra caution will be necessary.

[ahbarnett]

Looks great. I am in progress testing it out.
I had a question if type 3 with tol<1e-9 is requested, does it revert to upsampfac=2 ?
Best, Alex

[ahbarnett]

good - let's do it

[ahbarnett]

we should do it in setpts where nf is computed already for FFT sizes. Then choose 2.0 if FFT size will be small.

[ahbarnett]

Postpone to another PR

[ahbarnett]

Don't forget to explain in 1-2 lines in CHANGELOG.
Also rerun FFT=DUCC and see if need to add #ifdef
THanks! Alex