Para-Pred: Addressing Heterogeneity for City-Wide Indoor Status Estimation in On-Demand Delivery