[tune] use default anonymous metric _metric if at least a mode is set

[krfricke]

Why are these changes needed?

Ray Tune already supports an anonymous metric if e.g. tune.report(0.5) is called or a scalar value is returned in the (function) trainable. However, it is currently not used automatically, i.e. users will have to specify tune.run(train, mode="max", metric="_metric") to use this metric in the searchers, schedulers, progress reporting, and experiment analysis.

This PR changes this behavior so that the default _metric is used when no other metric has been specified but a mode has been set. Thus, the API reduces to tune.run(train, mode="max").

I think it's sensible to require setting a mode, as we cannot assume if we're optimizing for higher or lower values. Other libraries make this assumption, or default to one direction. Also, by explicitly setting a mode but not a metric, we can assume that the user actually wants to optimize an anonymous metric.

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/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[richardliaw]

looks like a kosher change though tests are failing?

Mainly comments regarding docstring.

[krfricke]

Failing test is test_trial_scheduler_pbt and the error seems unrelated to this PR (occurs in other PRs, too)