consider a separate send queue for PONG

[slingamn]

<JAA> In most non-trivial implementations I've seen, 'sending' actually adds to a queue, and the real sender throttles everything, but PONGs bypass the queue.
<slingamn> we don't have any special casing for PONG
<JAA> I'm not sure I've ever seen anything else bypass, although I bet there are exceptions.
<JAA> So if you build enough of a queue, you'll get disconnected with a ping timeout?

This is unlikely to come up in practice, but it's a theoretical possibility, We could add a second pwrite queue that just processes PONG. (select fairness would sort it out unless we get very unlucky; we could also deterministically favor the PONG queue by first doing a nonblocking select from the PONG queue, then a blocking select from both queues.)