Allgather bruck

[juntangc]

Adding a new algorithm for mpi collective allgather and reduce.

The allgather method is based on Bruck's original paper "Efficient Algorithms for All-to-all Communications in Multiport Message-Passing Systems" and a extension of current implementation in Open MPI (ompi_coll_base_allgather_intra_bruck). The new method use a fanout of k to allow multiple messages simultaneously send and receive on each node to reduce the number of communication steps required to complete the allgather operation (from log2(n) to logk(n)).

The new reduce method extends from recursivedoubling method and operations on a knomial tree instead of a binomial tree.