Two Failures of Self-Consistency in the Multi-Step Reasoning of LLMs