Avoid numerical overflow with overlapping atoms

Avoid numerical overflow with overlapping atoms

The verlet kernels did not allow overlapping atoms, even if they were
not interacting (in contrast to the group kernels). Fixed by clamping
the interaction distance so it can not become smaller than ~6e-4
in single and ~1e-18 in double, and when this number is later
multiplied by zero parameters it will not influence forces. The
clamping should never affect normal interactions; we would previously
crash for distances that were this small.
On Haswell, RF and PME kernels get 3% and 1% slower, respectively.
On CUDA, RF and PME kernels get 1% and 2% faster, respectively.

Fixes #1958.

Change-Id: I83b88f0e9ca34dc151a8b907f334a95a1a4301cc