Toward Explainable Offline RL: Analyzing Representations in Intrinsically Motivated Decision Transformers