Gate NUMA workaround on torch.numa presence, not torch>=2.10
The NUMA auto-detect that loads CUDA in the parent exists on torch 2.9 too
(torch.numa was introduced in 2.9), so gating the workaround on
is_torch_version(">=", "2.10") left notebook_launcher broken on torch 2.9.x.
Gate on `hasattr(torch, "numa")` instead, which matches the actual import.
