Benchmarking should not impact perf until timer is read

[cypof]

For timing operations in a net, it should be possible to add events to a GPU stream without synchronizing, to have a lower impact on performance and more accuracy. Synchronizing is only required when reading the timer, e.g. once at the end of an iteration instead of at each layer.

[shelhamer]

Looks good to me. Thanks Cyprien!

[shelhamer]

@flx42 could you comment on whether this resolves #4453? I see that you took another approach there with two passes, but is that necessary/is the overhead that significant?

[ujsyehao]

@shelhamer It is a big overhead.