Optimizing Seasonal-to-Decadal Analog Forecasts with a Learned Spatially-Weighted Mask