* Timeit now computes the median time from a select number of trial runs that
  are synced after each iteration
* The median time is then used to compute how many batches of the functions to
  run based on some minimum time
* It now runs batch of trial runs and then averages the times from each batch

* Synchronize the ArrayFire stream when cusolver getHandle is done
* The stream form of cusolver require using cudaDeviceSynchronize and
cudaStreamSynchronize all over the place
* Even then, getrs in solve_lu failed on any stream other than 0
* patch available at: https://gist.github.com/shehzan10/414c3d04a40e7c4a03ed3c2e1b9072e7