Split up Verlet SIMD kernels for faster compilation

Split up Verlet SIMD kernels for faster compilation

Each kernel is now defined in a seperate compilation unit, and the
files for the kernel functions, the kernel dispatcher function, and
their definitions are all generated by Python script. See its comments
for details. The generation is done by developers and the results
stored in the repo, to avoid CMake-time dependencies. The source for
the file generation is separated from the results of the generation.

On a 48-core AMD machine using GCC 4.7 for an SSE2 build (i.e. 4xN
Verlet kernels), the time for "make mdrun -j" went from 73 s to 34 s.

On a more recent 8-core Intel machine using GCC 4.6.4 for AVX_256
build (both 2xNN and 4xN), the time for "make mdrun -j" went from
81 s to 63 s.

On 16 cores of a POWER7 BlueGene/Q front-end node, time to compile
release-4-6 stays about the same (around 160s).

Also unwound some include file dependencies and removed repeated
definitions of some supporting structures and functions.

Change-Id: I0da1faf351defbe68d5ca43febcabddd93e21f0d