Structure-Preserving Margin Distribution Learning for High-Order Tensor Data with Low-Rank Decomposition