Using MPI_SendRecv instead of parity send/recv

[manaschai]

In send_recv routine, we can use MPI_SendRecv to simplify (and perhaps speedup) the code performance by get rid of the 2-phase send/recv based on node parity.

However, using MPI_SendRecv makes the exact size of incoming message (to be fill in rbuffer) unknown until the actual arrival of the message. Therefore, rbuffer needs to be large enough for the incoming message. From what I understand, we just make rbuffer sufficiently large (must be larger than the incoming message).

The implementation of send_recv is currently used in mk_pqeq branch c7452b8.