Address current problems in 3.10

Where the session is part-closed, these can throw NPE due to getRemoteAddress()
returning null. This sledge-hammers the problem away.

Addressing DIRMINA-1012

Do not synchronize on isClosed()

There is a deadlock while reading the state variable if close() is running.

This switches the state to be volatile instead - it's only written to inside
the lengthy close() lock, so this should be reasonable.

Openfire 3.10.1 Release Candidate

By closing a session when the new MINA inputClosed() handler is triggered, we run
the risk of sending data to the peer. As that peer is known to be dead, this is
pointless (and potentially dangerous - deadlocks have been observed that are likely
related to this scenario).

To prevent sending data during session closure, the close() method has been
overloaded with an argument that indicates if the peer is known to be dead. When
set, its implementation will not attempt to send data.

Currently, only stanzas containing errors of type cancel will cause the
occupant to be removed, however this doesn't appear to catch and remove many
occupants in practise.

This patch therefore removes occupants from the service on any message or
presence error stanza, irrespective of the error type. IQ stanzas do not
trigger the same response, as these may be due to lack of client support
rather than a vanished occupant.

The close listeners should not be called from within a synchronized block.

When a connection is closed, several events can be involved (and can be triggered).
Some of these events will attempt to close the connection (if it hadn't been already).
This, at best, least to multiple invocations of close-listeners. At worst, a loop of
close-calls is created.

This commit replaces the two-way boolean that guards closure (isClosed) with a
three-way guard. Now, a distinction is made between between a connection that is
closed, and one that is closing.

Additionally, some null pointer guards have been added, as I've seen those pop up in
my local logs during development.

MINA 2.0.8 introduces support for half-duplex close of TCP channels (see DIRMINA-785).
As part of this change, IoHandlers got a new method. The default behavior of the
IoHandlerAdapter closes the MINA session when this handler is triggered. However,
without an Openfire-specific extension (in ConnectionHandler), this fails to close
Openfire-maintained resources. As a result, CPUs could start to spin in NIO code,
resulting in 100% CPU cycles.

To fix this issue, ConnectionHandler now overrides the default functionality and
triggers a normal closure of the session (both in Openfire as well as MINA context).

Additionally, to prevent the CPU spin, MINA sessions need to be closed synchronously
(as opposed to the async closure Openfire had up until now). I cannot quite explain
the need for this change, other than a suspision of race conditions in the async
setup.

OF-909: Include ack in BOSH response body element

Fixes alternate (synchronous) write to the HTTP response output stream.

Per spec, the ack attribute in the response <body/> allows a BOSH client
to verify that sent stanzas with the corresponding rid(s) have been
successfully received by the server.

Openfire 3.10.0 GA Release!

Mark Openfire 3.10.0 Release Candidate

Change OF-888 fix to be based on error

OF-888 is presumed to be a recursion due to an attempt to bounce a bounce, in
other words it's failing to handle a double-bounce.

It appears to be caused when routing a stanza from a MUC fails, and for some
reason I don't yet understand, the routing of the bounce to the originating
MUC source also fails.

The test for the double bounce is only present for the IQ case, but this fails
in any case because it tests only for the IQ case, and uses a test which checks
the IQ's symbolic type; this is not actually set by Packet.setError(), so does
not trigger.

Tom's fix inserted a sentinel into the original failing stanza, however a new
stanza is created for the bounce, which will not contain the sentinel; therefore
the fix will not protect from a recursion.

Therefore this patch:

1) Removes Tom's fix for OF-888.

2) Tests for the message and presence error types.

3) Tests for the presence of a stanza error (though this is a warning).

4) Explicitly sets the stanza error type.

OF-901 update install4j windows bundled JRE to 1.7.0u76

OF-895 Update bundled PostgreSQL driver to 9.4-1201 JDBC4 (build 1201)

Note: this is java 1.6+ compat