Fix non determinism in backgrounds

Marginalized stats files have become nondeterministic on background runs as of !28 (merged). The fars are marginally different.
This MR happens to be just after our current external review, so it hasn't affected any test results, but it needs to be fixed or reverted before the following review.

I've run the following tests:
3 ~2h runs of f60408d2 in /fred/oz016/tdavies/projects/testing/run_fix_sampling_bg_short
3 ~2h runs of c065449f in /fred/oz016/tdavies/projects/testing/run_rerun_yapf_bg_short
The former is at !28 (merged), and most values in marginalized stats vary by relatively small amounts between runs.
The latter is just before it, and the marginalized stats are completely identical between runs.

There's already an issue with FARs in zerolags, described in #11 (prior to !28 (merged))
@andrewmichael.gozzard has been investigating, but some tests may have been confounded by this issue. It's also possible that !28 (merged) didn't introduce new nondeterminism, and only exacerbated the one causing #11. Even if that turns out to be the case, we should still revert it if it hasn't already been fixed prior to the next review.