Extending the WILDS Benchmark for Unsupervised Adaptation