CAD: Disaggregating Core Attention for Efficient Long-Context LLM Training (hao-ai-lab.github.io)