Previous commit exposes a bug in OpenMPI, switch algorithm for bcast to avoid it
The bug is that OpenMPI performs selection of algorithm based on the count on messages.
Or this count can be different from one process to another for bcast.
For example this test has a sender sending one element of 4096 bytes, and receivers receiving 1024 elements of 4 bytes
Yes, this is legal in MPI (but stupid ... but legal).
OpenMPI collective algorithm selector selects a different algorithm for these processes, thus deadlocking
