Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning