[Tune] Synchronous Mode for PBT

[amogkam]

Why are these changes needed?

Enable:

If True, will use synchronous implementation
            of PBT. Perturbations will occur only after all trials are
            synced at the same time_attr every perturbation_interval.
            Defaults to False. See Appendix A.1 here

Related issue number

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/latest/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failure rates at https://ray-travis-tracker.herokuapp.com/.
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested (please justify below)

[krfricke]

Just a question on the exploitation in synchronous mode, other than that it looks fine.

[krfricke]

Doesn't the exploiting trial still need to restore the state from the (new state) last checkpoint? I.e. why is there a distinction between synchronous and non-synchronous mode here, and why is on_checkpoint called in sync mode?

[richardliaw]

^ maybe leave a note about this

[amogkam]

Added a comment. If the trial is paused, we don't want to start it and then pause it again. We can just do on_checkpoint to update the most recent checkpoint, and when it gets started again, it will use the new checkpoint.