the current wait queue implementation isn't very good --- it's essentially a spinlock around a Vec. this is similar to what crossbeam uses, but worse :)

the downside of this approach is that when there's a large number of senders, there's severe tail latency because the vec has to be reallocated. also, memory usage doesn't decrease when the queue drains.

note this violin plot: when there are a lot of waiters, the tail latency is quite bad: