Previously, there were (basically) three states... active, closed,
and writeclosed where writeclosed indicated that a channel close
message was sent, no more writes are allowed, and it's waiting for
the close to be acked.

This adds a readclosed state where we have received a close request
from the remote (and acked it), but have not deactivated the channel
ourselves.

This fixes a race condition where we could select a channel, then,
when we write data to it, we receive a close and the channel becomes
inactive so a new active channel is selected (effectively at random).

This is a bit suboptimal though since we still don't get a
notification when a channel is closed by the remote.  Instead, if
we use more than one channel, we need to periodically make the
channel active for writes and explicitly check if it's closed.