Made some changes to the benchmark example... let me know what you think.

* Numpy does element-wise operations by default.  Updated the cpu
  operation to use pure numpy.

* Eliminated the loop which is not necessary to demonstrate
  parallelism on array operations.

* Made the number of workers explicit rather than gpu chosen,
  through local_size variable passed to kernel execution.

* Increased to ~8 million data points to more clearly demonstrate
  the difference between cpu and gpu based computations.

(integers work if they're intended as substrings of dev/plat names)