FilterEmpty yields unexpected behaviour #12
Description
Our training pipeline silently filter samples where the real-space neighborlist is empty. This is unexpected since we deal with long-range models: There are cases, for instance separating dimers, where there are no "short range neighbours". We should just drop this filter, as it is not required. However, we should check that our upstream dependencies (jax-pme in particular) don't unexpectedly fail for empty neighborlists.