Concentration Inequalities for Exchangeable Tensors and Matrix-valued Data