improved the nbnxn buffer size estimate with GPUs
The nbnxn Verlet buffer estimate now takes into account that
constrained atoms rotate, and don't move linearly, around the atom
they are constrained to. This significantly lower the buffer size
estimate for long neighborlist life times (as used with GPUs).
The buffer for most CPU runs is not affected (significantly).
Because of the smaller buffer, mdrun now uses smaller list increase
limits for increasing nstlist when using GPUs. This improves
performance.
Also activated and tested the virtual site effective mass calculation
(vsites were ignored in the drift calculation).
Change-Id: I2cb349f483610eabcc97bfbc23d17f189dec19d6