GPU memory monitor fails to log allocation.

[trivialfis]

When setting verbosity to 3 the memory logger should log all allocation and deallocation on GPU. But I encountered the deallocatation error when input is cudf, saying that the pointer in SpasePage::data is never allocated, which is obviously not true.

This is a reminder for myself. Also @RAMitchell might have some inputs here.

def test_cudf():
    X = np.random.randn(kRows, kCols)
    y = np.random.randn(kRows)

    df = pd.DataFrame(X)

    df = cudf.from_pandas(df)

    dtrain = xgb.DMatrix(df, label=y)
    bst = xgb.train({'tree_method': 'gpu_hist', 'verbosity': 2},
                    dtrain=dtrain,
                    evals=[(dtrain, 'X')],
                    num_boost_round=2)

    print('Pickle')
    import pickle
    with open('model.bst', 'bw') as fd:
        pickle.dump(bst, fd)

    with open('model.bst', 'br') as fd:
        bst: xgb.Booster = pickle.load(fd)
        dtest = xgb.DMatrix(df, label=y)
        bst.set_param({'gpu_id': '0'})
        bst.predict(dtest)

[trivialfis]

Found the cause ... verbosity is not effective during DMatrix construction.