Reassembling Consistent-Complementary Constraints in Triplet Network for Multi-view Learning of Medical Images