Do Unlearning Methods Remove Information from Language Model Weights?