Looking at the nwchem.py script, which does not come from the NWChem developer's team it looks like ASE only gets gradients from NWChem. So, ASE itself does the NEB calculation using the gradients I assume. You might want to try doing a very coarse convergence of the NEB to see if that does converge. If it does, then the convergence criteria in the nwchem.py might not be tight enough.