On Domain Generalization Datasets as Proxy Benchmarks for Causal Representation Learning