Concentration of Non-Isotropic Random Tensors with Applications to Learning and Empirical Risk Minimization