Effective Multimodal Reinforcement Learning with Modality Alignment and Importance Enhancement