switched to nanmeans for t-tests

[HeikoSchuett]

For small datasets individual bootstrap evaluations can become nan
In these cases we still want to get p-values from the tests, i.e. use nanmean instead of mean.

[HeikoSchuett]

@ilogue could you have a look what breaks these tests? I cannot reproduce the error locally and did not change the odd even split code at all

[JasperVanDenBosch]

@ilogue could you have a look what breaks these tests? I cannot reproduce the error locally and did not change the odd even split code at all

Yeah it's definitely not related to these changes. I'm investigating in #336. I suspect it's just that one of our dependencies has changed its behaviour.