Improve inclusive/exclusive_scan() performance

[kylelutz]

Improve the performance of inclusive_scan() and exclusive_scan(). Scan multiple values per thread. This will also improve the performance of partial_sum().

[banche]

I will give a look at this week, as far as I remember a quite efficient scan can be done in 3 distinct steps : Up-Sweep, Down-Sweep and then compute the final result ( or something like that).

Hope I can improve the performances!