Near-Optimal Multi-Perturbation Experimental Design for Causal Structure Learning