Add check for cuda lazy init (#80912)
Validate that no CUDA calls are made during `import torch` call, by
importing torch and limited visible devices to non-existing device
Should prevent regressions like ones reported in https://github.com/pytorch/pytorch/issues/80876
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80912
Approved by: https://github.com/ngimel, https://github.com/atalman
