Add high-level API Scatter for splitting 1D array

[ykkan]

Following the idea of Gather, a high-level API Scatter for Scatter! is provided. With this API, users can easily scatter a 1D array without handling VBuffer or UBuffer.
For example:

MPI.Init()
comm = MPI.COMM_WORLD
rank = MPI.Comm_rank(comm)
nprocs = MPI.Comm_size(comm)


arr = nothing
if rank == 0
  arr = [1:10...]
end

local_arr = MPI.Scatter(data, comm)

print("Hello world, I am rank $(rank) of $(nprocs), $(local_arr)\n")

Output (nprocs=3):

Hello world, I am rank 0 of 3, [1, 2, 3, 4]
Hello world, I am rank 2 of 3, [8, 9, 10]
Hello world, I am rank 1 of 3, [5, 6, 7]

[vchuravy]

Why is one of these Bcast and the other bcast?

[ykkan]

If I remember correctly, elm_t is not a isbits type and can only be broadcasted with bcast. My idea was to use Bcast whenever possible, as its overhead should be less than bcast(?). Of course, using bcast for both elm_t and arr_len should work. I think my idea behind is rather a matter of style. What do you think?