The retire worker is kicked for each fence, either the normal way
by signaling the fence from the event completion interrupt or by
the recover worker if the GPU got stuck. Moving the RPM put into
the retire worker allows us to have it in a single place for
both cases.

This also shaves off quite a bit of the CPU time spent in hardirq
context, as arming the autosuspend timer when the RPM refcount
drops to 0 is a relatively costly operation.
Tested-by: Russell King <rmk+kernel@arm.linux.org.uk>
Acked-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>