Uniform treatment of forces on "single-particle" vs. normal atom groups

[jhenin]

To properly handle fit gradients for all CVCs in an efficient way, we need to know whether atom groups are involved as their COG/COM, or if the cvc depends on individual coordinates in a more complicated way.

A more general approach is to store a vector of weights, providing for all linear combinations of coordinates (or a weight matrix for vector cvcs).

In those cases, we need to store per-group gradients, instead of per-atom.

[jhenin]

Side issue to this: treat dummy atoms in a unified way with regular atoms.

[giacomofiorin]

This is an easy fix after completing #61