Gone but Not Forgotten: Improved Benchmarks for Machine Unlearning