Hierarchical Neural Additive Models for Interpretable Demand Forecasts