STREAM Benchmark in CUDA C++

Variant of the STREAM benchmark written in CUDA C++, based on work by Massimiliano Fatica (NVIDIA).

Three models are available:

1. default: the original GPU STREAM using HBM.
2. -DUSE_HOST: use managed memory pointer allocated on the host, cuda runtime may make automatic page migration.
3. -DZERO_COPY: use coherent access in unified memory architecture, directly access host memory without data movement.

To Do: add output validation.