Towards Representation Learning for Weighting Problems in Design-Based Causal Inference