Theoretically Grounded Pruning of Large Ground Sets for Constrained, Discrete Optimization