When we close a connection via connection_close_immediately, we kill
its events immediately. But if it had been blocked on bandwidth
read/write, we could try to re-add its (nonexistent) events later
from connection_bucket_refill -- if we got to that callback before
we swept the marked connections.

Fixes bug 24167.  Fortunately, this hasn't been a crash bug since we
introduced connection_check_event in 0.2.9.10, and backported it.

This is a bugfix on commit 89d42291, I believe, which
appeared in Tor 0.1.0.1-rc.