Interpretable Reinforcement Learning via Neural Additive Models for Inventory Management