Fast Slate Policy Optimization: Going Beyond Plackett-Luce