Provable unlearning in topic modeling and downstream tasks