Use cudf.concat explicitly. (#4918)

* Use `cudf.concat` explicitly.

* Add test.

tests/python-gpu/test_gpu_with_dask.py

@@ -0,0 +1,41 @@
+import sys
+import pytest
+
+if sys.platform.startswith("win"):
+    pytest.skip("Skipping dask tests on Windows", allow_module_level=True)
+
+try:
+    from distributed.utils_test import client, loop, cluster_fixture
+    import dask.dataframe as dd
+    from xgboost import dask as dxgb
+    import cudf
+except ImportError:
+    client = None
+    loop = None
+    cluster_fixture = None
+    pass
+
+sys.path.append("tests/python")
+from test_with_dask import generate_array
+import testing as tm
+
+
+@pytest.mark.skipif(**tm.no_dask())
+@pytest.mark.skipif(**tm.no_cudf())
+def test_dask_dataframe(client):
+    X, y = generate_array()
+
+    X = dd.from_dask_array(X)
+    y = dd.from_dask_array(y)
+
+    X = X.map_partitions(cudf.from_pandas)
+    y = y.map_partitions(cudf.from_pandas)
+
+    dtrain = dxgb.DaskDMatrix(client, X, y)
+    out = dxgb.train(client, {'tree_method': 'gpu_hist'},
+                     dtrain=dtrain,
+                     evals=[(dtrain, 'X')],
+                     num_boost_round=2)
+
+    assert isinstance(out['booster'], dxgb.Booster)
+    assert len(out['history']['X']['rmse']) == 2