EVADE: LLM-Based Explanation Generation and Validation for Error Detection in NLI