This is basically a bug to break apart bug 1202 into more manageable pieces.

Supposed the ORB is running nested event loops to some connection (say the
connections is flow controlled), now suppose the connection is closed by the
peer (for example it crashes.)  The ORB should notify all the queued messages on
that connection about the closure, so the nested event loops can exit.

However, several places in the ORB close the connection and then expect the
destructor of the Transport for notify the queued messages, however, the
destructor won't get called because there are still references to the transport
(in the nested event loops!)

The Nested_Upcall_Crashes test can reproduce this problem (sometimes), I expect
I'll be able to submit patches shortly.