Matrix Vector Kernel Parallelism

[MrBurmark]

The POLYBENCH_MVT, POLYBENCH_GESUMMV, POLYBENCH_GEMVER, and POLYBENCH_ATAX kernels (and potentially more) may be able to express more parallelism when computing dot products by using level block reductions or atomics.