More robust t-tests

[nilshoffmann]

This PR proposes to add an additional parameter to the compare_means test methods for cases where groups have too few members or when calculation throws any other errors, but should be kept in the same workflow. The value p.error.default.value is currently exposed with a default of 0, but could also be changed to a boolean parameter p.error.recover = FALSE as default to keep the standard behaviour or p.error.recover = TRUE to use 0 as the fallback p-value if the test method fail for an input.