ghostAtoms

When N3L, all ghost atoms will get a force. This force is added to the corresponding original atom
when all the forces are calculated. Therefore the ghost atom index and not the tag should be used, because
the ghost atom incidies appearently do not correspond to the original index, or tag. 

If newton_pair is not on, the force is not added to the originals? Then then sum k > j wont work,
because here there are also ghost atoms. So N3L must be on for this to work.
But what about the pairs? N3L should be incorrect when I count all pairs during training. Do I have to use N3L for both pairs and triplets?
Does it automatically work if I do the N3L test when I sample LAMMPS data?

The force on atom i is the gradient of the total energy from all neighbouring atoms. But the energy on atom i due to the presence of atom j inside
atom i's cutoff sphere is shared between the two. Therefore the total energy on atom i is half the energy. One must use half the energy
to get the correct force? But then the NN should maybe just count half of them? This must be considered when not training against forces. 
If the NN counts all pairs and get twice the energy as in lammps, then it learns that the energy is not shared. The energy value
directly influences the force obtained when differentiating, resulting in wrong force. If the NN counts all pairs and learn to yield
half the energy, then it learns that the energy is shared, but the resulting lower force is distributed on too many particles, 
resulting in too small forces. The correct way (???) is to make the NN learn that the energy is shared, in addition to distribute
the resulting force on just half the number of particles at a time.