added Verlet scheme and NxN non-bonded functionality

added Verlet scheme and NxN non-bonded functionality

This commit implements a new "Verlet" cutoff scheme which uses
a exact cut-offs and standard Verlet lists with an automatically
calculated buffer.

The Verlet code-path supports full multi-level heterogeneous
parallelization using MPI/thread-MPI, OpenMP multi-threading,
and GPU acceleration for the non-bonded calculations.

The non-bonded calculations with the Verlet scheme support highly
optimized CPU SIMD acceleration using SSE/AVX and GPU acceleration using
NVIDIA CUDA. The CPU kernels have been tested on and optimized for most
x86 architectures including recent ones like Intel Sandy/Ivy-Bridge and
AMD Bulldozer. The CUDA GPU kernels support hardware of compute
capability 2.0 and above and are optimized for both Fermi and Kepler
architectures.

The new search code has been added in nbnxn_search.c, new non-bonded
kernels in nbnxn_kernels and nbnxn_cuda:
- plain-C kernels: reference CPU implementation and reference GPU
  (emulation)
- x86 128- and 256-bit SIMD kernels (SSE2, SSE4.1, AVX_128, AVX_256
  intrinsics)
- CUDA (two versions: for recent and legacy toolkit/drivers)

This commit also implements some additional optimizations targeting
performance:
- SSE acceleration for dihedrals;
- automated PP/PME load balancing called "PME tuning" which optimizes
  the electrostatics cut-off to improve load balance between CPU and GPU
  or separate PP and PME processes;
- hardware detection and automated run-configuration selection.

Change-Id: I3e1a15331c174265ec086565b978ffd079df2aaa