Update ddp+dynamo benchmark arguments (#1222)
Summary:
Pull Request resolved: https://github.com/pytorch/benchmark/pull/1222
* Reduce default timeout to 2 hours; afaik, all of these pass well
within 2 hours unless there are configuration issues (e.g. bad nodes,
etc.). Reduced timeout means that we avoid blocking the cluster while
waiting for a dead job, which also means that _our_ experiment will also
probably finish faster because it's not blocked on dead jobs.
* Add `--exclude` argument which can be used to avoid bad nodes.
Test Plan: Imported from OSS
Reviewed By: wconstab
Differential Revision: D39952268
Pulled By: davidberard98
fbshipit-source-id: cff2526566aa46ca094e4db6aea4422b44bc710f
