[tune] Checkpoint according to nested metric

[ghost]

Why are these changes needed?

In tune, checkpoint according to a metric nested in the reported result.

This is useful when optimize a metric of evaluation in rllib.

My solution is similar to #14375

Related issue number

Closes #14377

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 :(

[ghost]

I noticed that the unit test failed, but I don't quite understand error message.
At the end of log, it points out that 79 test pass and 1 test failed. But there are multiple lines indicating test failure.
Can anybody help me locating the error? I can try to solve it if I know exactly which test failed.

[krfricke]

The Travis CI test is currently the one we use for tracking - sometimes other tests are flaky and can lead to failing CI. The changes from this PR apparently did not introduce any failures, so we should be good to merge

[ghost]

Ah, got it. Thanks a lot!