Hard Constraints, Smooth Gradients: Learning Feasible Inventory Policies via Differentiable Projection