Fix non-determinism in FAR calculation

FARs are non deterministic as far back as !20 (merged), and likely earlier. They haven't affected injection tests, but appear significant enough (+-0.1) to affect marginal triggers. The issues are sporadic, so we may have just been lucky so far.

There are streches (usually several triggers or ~20s long, sometimes much longer) of large differences in FAR (~0.1).

!20 (merged) had other types as well (long streches of very small differences). I haven't seen them after !20 (merged), and other more sporadic issues are probably the result of changes to finalsink clustering.

Either way, we should start by investigating our KNN and FAR calculations.