This is already partially supported but never properly tested. So the only possible way to use it is calling `numpy.ndarray.flatten` with `base_margin` before passing it into XGBoost. This PR adds proper support
for most of the data types along with tests.