This optimisation merges successive calls of u_ghost to increase the number of communications before waiting.