Optimal bounds for $\ell_p$ sensitivity sampling via $\ell_2$ augmentation