FP16 support

[kuenishi]

At least we know with FP16 model Communicator's bcast_data does not work.

diff --git a/tests/chainermn_tests/communicator_tests/test_communicator.py b/tests/chainermn_tests/communicator_tests/test_communicator.py
index a0ff350..f03fa5d 100644
--- a/tests/chainermn_tests/communicator_tests/test_communicator.py
+++ b/tests/chainermn_tests/communicator_tests/test_communicator.py
@@ -242,6 +242,12 @@ def test_communicator_cpu(param):
     check_send_recv(param, False)
     check_collective_communication(param, False)
 
+    communicator = create_communicator(param, False)
+    communicator.mpi_comm.barrier()
+    x = np.random.rand(1, 3).astype(np.float16)
+    x = communicator.mpi_comm.Bcast(x)
+    print(x)
+
 
 @pytest.mark.parametrize('param', gpu_params)
 @chainer.testing.attr.gpu

• non-cuda aware communicator does not support float16 model as MPI does not support float16.

[shu65]

It is solved by #271 when using pure_nccl