follow test8, 1. 新增torch后端_torch_backend.py，numpy/cupy输入自动转换、complex64/128精度切换与numpy-like适配层; 2. 新增pyqcd/parallel MPI并行模块（plan_parallel按N*a=n*b公式规划、run_parallel_pipeline元任务round-robin调度+GPU绑定）; 3. 重构IO层为h5py读写（save/load_tensor_h5，_load_any兼容旧.npy/.npz产物）; 4. 新增docs/analy_torch_parallel_20260816分析文档并入库PDF; 5. 修正refer/git-rep/EasyDistillation与PyQUDA的.gitignore并入库pure文档PDF; 6. AGENTS.md同步torch后端/h5 IO/MPI并行约定与概要命令; [opencode].