Improve load imbalance.

[rainwoodman]

HiddenValley3 (HV3-10240-Gaussian) simulation pushed to z=0.5, but the load imbalance is already 5.6. This is a waste of memory and slows down the simulation too.

We can merge particles that are known to have already formed tiny bound structures (such that they are unlikely significantly disrupted by tidal force). This will significantly reduce the load imbalance.

The code already support variable particle mass (@adrianbayer's neutrino work) via the variable 'additional' mass column; however, for this we may want to keep the exact number of primordial particles a simulation particle represents by adding a new int8 or int16 field.

We also need to modify FOF finder to take this into account.

In the snapshot writer, we may want to expand a merged particle into several primordial particles at the same location, or to spead the primordial particles by a standard profile -- to preserve compatibility with existing tools that assume uniform particle masses.