optimize MPI_Sendrecv if a process has to exchange with himself only

This helps avoiding some lockings of lazy network model when one process waits for itself