SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities