Exposing LLM-Generated Logical Flaws in Reasoning via Automated Theorem Proving