Fitting tests with random noise are flakey

[dnadlinger]

See e.g. https://app.circleci.com/pipelines/github/OxfordIonTrapGroup/ndscan/1293/workflows/212100d7-8319-49af-9884-710ff675623d/jobs/2605.

We should probably hard-code the seed (and improve the oitg heuristics, but that's not something we want to depend on here).