Infrastructure for new group kernels, and new C kernels

Infrastructure for new group kernels, and new C kernels

This patch implements a new selection mechanism for nonbonded
kernels, as well as modifications of the neighborsearching to allow
long-range lists to be used every step (for rlist!=rcoul!=rvdw),
which is necessary for PP-PME load balancing. We can also evaluate
long-range interactions at steps that are arbitrary divisors of
nstlist. Kernels are now selected from text strings describing
desired options, which makes it much easier to have different
kernels for different architectures and gradually add kernels for
commonly used cases without implementing all combinations of
features. There is also a new kernel generator and flexible
preprocessor written in python. This is only executed by developers,
and the generated kernels are therefor included in the git tree
after discussions. The actual kernels are now written in plain C
with slightly more advanced preprocessor options allowing loops. The
listed (1,4) nonbonded interactions now use their own code rather
than executing a complete kernel call for each pair, which saves
a lot of overhead. The old c kernels have been adapted to continue
working, but addres kernels are disabled by this patch. Subsequent
patches will start adding back the kernels (one per architecture),
and also the addres kernels. The group kernels now support arbitrary
combinations of switch and shift for most interactions (but the
might be run in the generic kernel), and there are new kernels
specifically designed for Ewald direct space interactions as well as
force-only kernels.

Change-Id: Iab142bcc2e6d914ca127b4caa6a07b604549cb13