xJailbreak: Representation Space Guided Reinforcement Learning for Interpretable LLM Jailbreaking