Add test for out-of-bounds Tensor access on GPU (#39211)
Since CUDA context can not recover safely from on-device assert, use `torch.multiprocessing.spawn` to execute a method in another context and verify that it raises unrecoverable error.
As those types of tests are pretty slow (6 seconds on powerful linux box with one GPU) run it only in the slow shard.
Closes https://github.com/pytorch/pytorch/issues/38944
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39211
Approved by: https://github.com/ezyang
