From MNIST to ImageNet: Understanding the Scalability Boundaries of Differentiable Logic Gate Networks