fix: auto-download dependency DLLs when worker fails to load on Windows

- worker.py: cuDNN scanner now adds all CUDA version subdirs (was only
  picking the first, missing CUDA 13 cuDNN when 12.x came first)
- nodes.py: _ensure_worker downloads dep zip from ModelScope on first
  DLL load failure, then retries; thread-safe with download lock;
  returns clear error if download is in progress
- lib_setup.py: _test_load_dll uses subprocess with full path scanning
  to avoid false positives from PyTorch in-memory CUDA DLLs
- auto_update.py: blocks on first-time library download instead of
  racing with user actions in background thread