Efficient Long-context Language Model Training by Core Attention Disaggregation