Unlocking General Long Chain-of-Thought Reasoning Capabilities of Large Language Models via Representation Engineering