The Limit Points of (Optimistic) Gradient Descent in Min-Max Optimization