Falling back from CUDA to CPU torch leaves torch unimportable on Windows #731

Split out of #723.

What goes wrong

After a CUDA torch install fails verification on Windows, the desktop app puts CPU torch back. The result cannot be imported at all: OSError: [WinError 127] ... c10_cuda.dll. From then on the backend cannot run separation on any device, and setup still reports success.

Why

The restore installs CPU torch over the CUDA one with pip install --ignore-installed --no-deps. That overwrites files both wheels share, but never removes files only the CUDA wheel has, such as c10_cuda.dll and torch_cuda.dll. On Windows, torch loads every DLL in torch/lib, so the leftover CUDA DLL from one torch version fails to link against the restored CPU DLLs from another.

Nothing checks that torch imports after the restore.

Constraints

• The torch that shipped in the bundle is known to work, and it is on disk before any CUDA attempt. Putting back exactly that needs no network.
• A crash or forced quit in the middle of a swap must be recoverable on the next launch.
• Installs that are already broken have no backup to restore from.