Landscape Surrogate: Learning Decision Losses for Mathematical Optimization Under Partial Information