Fix Hafnian accuracy issue

[GregoryMorse]

No description provided.

[Kolarovszki]

Seems good to me. In conclusion, the problem was that sqrt operated with doubles instead of long doubles, right?

[GregoryMorse]

Yes exactly, the pure double and pure long double modes were new. But for proper accuracy comparison to the other library, we needed the long double implementation completely proper. And for some reason gcc seems not to be standards compliant.

sqrt(long double) and sqrt(double) should simply work in fact according to: https://en.cppreference.com/w/cpp/numeric/math/sqrt

Perhaps in more current gcc versions this library issue is fixed, as it has been reported, as its much broader, we previously ran into it with integer rounding with rint vs lrint and such.