tf.space_to_batch_nd