Re-evaluate CUDA NVML and fork checks after torch 1.13 release

[awaelchli]

🚀 Reminder

This is a reminder that we added code in #14984 and #14631 that will potentially become obsolete depending on what 1.13 contains.

• If Extend torch.cuda.is_available() to attempt an NVML-based CUDA availability assessment when explicitly requested by the user pytorch/pytorch#85951 lands and makes it into 1.13, we're good and can remove the code when we drop support for 1.12. There is already a note about it in the source code.
• If Extend torch.cuda.is_available() to attempt an NVML-based CUDA availability assessment when explicitly requested by the user pytorch/pytorch#85951 gets pushed to 1.14, we might have to bump our internal checks.

cc @justusschock @awaelchli @speediedan @carmocca

[carmocca]

This can be closed now, right @awaelchli? The checks are for 2.0 now