Update benchmark results in CUB documentation

[sh1ng]

@dumerrill the library is awesome!

All documented results are out of date and there's no results for Pascal generation. Probably you already have that code but not publicly available.

[alliepiper]

Just to put this out there, many of the existing CUB tests include benchmarks that can be activated by passing in the --i [num_iterations] option to control the number of benchmark iterations. Many also support --n [num_values] to change the problem size.

That said, I'm working on moving Thrust and CUB to a unified benchmarking system, so stay tuned.

Updated the title to something more actionable.