Recent changes to res_pjsip_outbound_publish have introduced a
race condition at shutdown where an outbound publish may be shutdown
twice. In this case the first succeeds as a result of the unpublish.
In the second invocation since it's been unpublished a task is
queued to just destroy the client. This task holds no ref to the
publish and as a result the publish may be destroyed before the
task is run, causing a crash.

This explicit destruction task now holds a reference to the publish
to ensure it remains valid.

ASTERISK-26053 #close

Change-Id: I10789b98add3e50292ee3b33a55a1d9061cec94b