Run slow gradcheck on the newer G5 runner (#102496)
As slow gradcheck is slow (Thank you, Captain Obvious!), let's run it on the newer G5 runner to improve its TTS and avoid flaky timing out error such as https://github.com/pytorch/pytorch/actions/runs/5112059782/jobs/9190167924. AFAIK, there is no reason to keep running slow gradcheck on `linux.4xlarge.nvidia.gpu`
### Testing
* `1st` shard: `3h30m` → `4h`, The increase is probably due to https://github.com/pytorch/pytorch/pull/102380 in which the job's name switch from `gcc7` to `gcc9`. Does this invalidate the test time used to balance these shards?
* `2nd` shard: `4h35m` → `4h15m`
* `3rd` shard: `3h20m` → `1h20m`
* `4th` shard: `3h20m` → `2h10m`
* `14h45m` → `11h45m`, a total saving of `3h`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/102496
Approved by: https://github.com/malfet
