[todo] Investigate whether we should turn on NETKET_EXPERIMENTAL_FFT_AUTOCORRELATION by default

[PhilipVinc]

setting NETKET_EXPERIMENTAL_FFT_AUTOCORRELATION=1 makes sure that we use a proper autocorrelation time calculation based on FFT to estimate the autocorrelation time in the chains.

This is a bit more expensive that the naive algorithm for estimating the autocorrelation time that is used by default, which is sometimes wrong.

We never looked into how much more expensive NETKET_EXPERIMENTAL_FFT_AUTOCORRELATION is and whether we can switch.
I personally would prefer to switch over to more correct results, even if it comes at a small cost. but this should be benchmarked...

If someone wants to investigate this:

• we should benchmark how expensive it is to estimate expectation values with and without this flag (it can be turned on and off by doing nk.config.netket_experimental_fft_autocorrelation = True/False ) for very simple observables, for different number of samples.
• see how much performance is lost.