iriri

I was looking at the `SegQueue` implementation today and also noticed the somewhat unfortunate allocation behavior so I threw together [a branch](https://github.com/iriri/crossbeam/tree/cache-segqueue-block) where up to one block is cached for...

>If one were to limit the junkyard to be pointer size of the machine, you could simply use an atomic bitmask which might be simple enough to beat allocation. So...

Decided to revisit this after some time and added a rough [latency benchmark](https://github.com/iriri/crossbeam/blob/a85ff05ae75d508bd77ac40a7799ee1bfa5b0d5a/crossbeam-channel/benchmarks/send-latency.rs). There is some run-to-run variation (a lot, actually, in the case of `master`) but these results are...