Fix ELCIC nonneutral for CPU, disable for GPU

[reinaual]

Hopefully fixes #3001, disables combination described in #5408

This now finally re-enables ELCIC for non-neutral systems. It turns out, that also the not yet disabled metallic case was also affected by #3001.

• For neutral systems, the dipole correction is translational invariant. However, for non-neutral systems it's not. So one part was to measure the dipole moment with respect to the symmetry-plane of the system. This caused a constant force towards one of the plates (depending on a sign), which is now corrected.
• The force was not the gradient of the energy. Due to a mistake in the sign of the energy and a mess between lz with and without gap for the corrections, we had a jump in our energy.
• The background_energy-correction counted the energy correction twice.
• The analytic ImageSum used box_length instead of box_h.
• I've enabled non-neutral with non-metallic boundaries as well, as I don't see why this was forbidden. To the best of my knowledge it works exactly the same as the metallic case.
• Fixed an interface issue, where const_pot was silently ignored by the code.
• Disabled P3M-GPU with dielectric ELC, due to ELCIC forces wrong with P3M-GPU #5408