Fixed issue with combined k-eff estimator variance producing NaNs
Created by: nelsonag
This PR fixes issue #613 (closed). In that issue, @paulromano points out that sometimes (though probably very rarely) the combined k-eff variance estimator produces NaNs. I also mentioned that the floating point exceptions also happen (every time) when using MG mode with survival biasing enabled.
This bug definitely fixes the latter (MG mode + surv biasing), and was also written in a generic enough way to probably also fix the former problem.
The method was derived following the exact same process detailed in our current methodology (described in "Estimation and Interpretation of keff Confidence Intervals in MCNP" by Urbatsch et al), except two estimators were used instead of three in the reference. I did the derivations by hand, but as a sanity check I also used a sympy Jupyter notebook to repeat the calculations. The two methods produced perfectly equivalent formulas (phew).
That jupyter notebook is here. I can scan in the hand derivations too if you really want them, but you probably wont be able to read them.
This PR also adds a test of the MG-mode survival biasing