On the Self-Verification Limitations of Large Language Models on Reasoning and Planning Tasks