Probing the Safety Response Boundary of Large Language Models via Unsafe Decoding Path Generation