The serial version of this algorithm would have an outer loop, i, and an
inner loop, j.

We parallelize the i loop by thrust::transform(); for each i, it
"transforms" the i, the output being the value of the j loop for that i.
We then add across the i loop, by calling thrust::reduce().

Could add randomization of i's for possibly better load balance, as in
rthconv().

This could be speeded up by using a transform iterator with the reduce
operation.