Learning Linear Programs from Optimal Decisions