Certified $\ell_2$ Attribution Robustness via Uniformly Smoothed Attributions