Benchmarking General-Purpose In-Context Learning