TRDA-TS: Text reinforcement and regularized diffusion alignment for LLM-empowered multimodal time series forecasting