Refactor functions to use MPI_Allreduce

[hakonsbm]

Fixes #1925 by refactoring MPIManager::any_true() and RandomManager::check_rng_synchrony() to use MPI_Allreduce.

[heplesser]

Looks good, but I think there is a corner case for which I suggest a solution.

[hakonsbm]

@heplesser I didn't think of that corner case, thanks for pointing it out. However, I don't agree with moving the local non-equality check into the equal_cross_ranks() function. The equal_cross_ranks() function has one job: to check if a value is equal across all ranks. Local non-equality is a special case, which should be handled locally before checking if values are equal across all ranks. Another problem with handling local non-equality in the equal_cross_ranks() function by passing local_min and local_max values is what to do in cases where there is only a single value, like when checking the rank-synchronized RNGs.

In the case where there is a local non-equality, this has to be communicated to all ranks. I've taken your suggestion of using -inf instead of -1 to flag a local non-equality, which will become the result of values[ 1 ] after communication and in that way raise an exception in all MPI processes. Please have another look.

[heplesser]

@hakonsbm You got a point. My idea was to isolate the "-inf" magic entirely in the equal_cross_... function but then one should probably have an equal_across_vps() functions. Your solution is fine and I will approve.

[hakonsbm]

@mlober Friendly ping

[mlober]

Hi, sorry for the super late reply, I completely missed this. I looked through everything and I like your solution. Thanks for the improvement!