If you had a chain of packets being sent all at once, the latter
section of packets could be delayed, theoretically infinitely.
If queue1 was filled and enqueued for transmit, then queue2
had packets enqueued in it while queue1 was still transmitting,
we'd try to fill queue2 further rather than waiting for queue1's
transmit to finish.
However, in that case, we wouldn't check if we could transmit
queue2 again until the next packet. If the user application
was waiting for the response from something in queue2
before pushing more packets, it could hang indefinitely.
This also fixes a subtle bug where hitting the "not safe to try
to fit any more packets in this queue" limit would cause a
packet to drop, as it would be dequeued and then tossed.
Closes GH-42
0ce7326b83