`cuda._is_in_bad_fork`->`_C._cuda_isInBadFork` (#87317)
Former is always available, while later is only available if PyTorch compiled with CUDA And if it does, then
```
$ python -c "import torch;print(torch._C._cuda_isInBadFork == torch.cuda._is_in_bad_fork)"
True
```
Fixes https://github.com/pytorch/torchdynamo/issues/1709 ( at least the symptom)
cc @jansel @lezcano @fdrocha
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87317
Approved by: https://github.com/voznesenskym, https://github.com/albanD, https://github.com/soumith, https://github.com/jansel
