Add pytest workflow and the Ray mode.

[garrett4wade]

1. Add a pytest github workflow to run pytest -m "not slow" upon each pull request.
2. Support Ray mode. Quickstart commands with mode=ray will launch distributed jobs with ray. This mode assumes that the user has launched a ray cluster with ray start CLI. Launching ray clusters with srun is no longer supported because it increases the complexity.
3. Move the CUDA GAE test from legacy code to the current repo.
4. Add more comments in the scripts in the examples folder.
5. Fix the unnecessary nan/inf check when loading/saving checkpoints. Now save & load is as fast as before Loading parameters takes exceptionally long time. #18 .
6. Fix a generate bug when only the LM head resides on the last pipeline stage.

Future works:

• Change the doc correspondingly.

[nuzant]

LGTM