vchiq_arm: Avoid use of mutex in add_completion

Claiming the completion_mutex within add_completion did prevent some
messages appearing twice, but provokes a deadlock caused by vcsm using
vchiq within a page fault handler.

Revert the use of completion_mutex, and instead fix the original
problem using more memory barriers.

Signed-off-by: Phil Elwell <phil@raspberrypi.org>