Measuring Faithfulness of Chains of Thought by Unlearning Reasoning Steps