Self-Similarity Priors: Neural Collages as Differentiable Fractal Representations