Optimization: Only drain queue if it would otherwise be full

[ftschirpke]

After reading the reference implementation for a while regarding burst control and #45, I found that Tahir et al. actually do not drain "fluently" in the same way that we do. They actually only drain the queue if it otherwise would be full, and then use a "fluent" approach by calculating this drain basically as now - last_dequeue_time.
(See https://github.com/PhantomQueuePolicer/BCPQP/blob/ef563248f504d36fa303ee932e201004f945171b/testbed/machnet/src/apps/bcpqp/main.cc#L582)

[corite]

That's quite clever. With that we could reduce the number of compare_and_swaps to whenever the queue is full (compared to on every packet right now). That would of course need some changes in try_increment_counter which tend to be a bit finicky, but it could be worth it. I'll take a look.