Allreduce: Reduce then Bcast
Alltoall: "basic_linear" if data per proc < 3Kb, "otherwise pairwise".
Not yet implemented: "Bruck" for data per proc < 200b and comm size > 12
+ Alltoallv: flat tree, like ompi
Scatter: flat tree
* Add support for optimized collectives (Bcast is now binomial by default)
* Port smpirun and smpicc to OS X