CuEvent and performance benchmark in device time

[moon6pence]

Implement interface for using cuda events.

I did it long time ago in CUDA.jl package, so it will be simple task. https://github.com/moon6pence/CUDA.jl/commit/ab1aed855839cffd0f37da9cc03d69967b2a5e6a

And we can implement such @elapsed and tic/toc() stuff in GPU time. We can use CPU time functions but GPU time looks be more accurate and usually gives faster result.

[timholy]

Sounds good.