Adapting LLMs to Time Series Forecasting via Temporal Heterogeneity Modeling and Semantic Alignment