Adaptive importance sampling for heavy-tailed distributions via $\alpha$-divergence minimization