allow disabling the error checking the benchmark

[mkstoyanov]

The benchmark sanity error check compares the result from a forward and inverse transforms against the original input. However, the data is not reset between the repetitions, which will create more work (memcopy) and may skew the benchmark. Combined with many runs, e.g., 2000 runs with single precision in 128^3 problems, the error may exceed the tolerance. In addition, the error check uses another MPI_Allreduce(), which is not counted in the total timing but can still take a while.

• added a flag to skip the error checking in the benchamrk.