Automatic distributed-memory reduction wrap-up

Allows the automatic wrap-up of distributed-memory reduction patterns.

Assuming tasks have been correctly inserted on each contributing node, the following happens :

• if users want to specify how the reduction scheme happens ; then, they call starpu_mpi_reedux_data()
• if users trust StarPU to wrap reduction patterns ; then, reduction trees are submitted upon calling starpu_mpi_wait_for_all()
• if users actually use the reduced value ; then, reduction trees are submitted before processing the send requests associated with the use of the reduced value

mpi_redux_autowrapup.c example extends mpi_redux.c with the 3 behaviours.