[tune] Simplify FunctionRunner and unify logging behavior with Trainable

[gehring]

Describe the problem

Currently, the function runner is quite complex for a wrapper and involves hard coded sleeps which can cause results to skipped over when they occur too fast. This issue is meant as a feature request for a simpler function runner with logging following the same behavior as the Trainable class.

Related issues

#3834

[gehring]

@richardliaw I missed the change to which fields are checked in the consistency check.

Not a pressing issue but I do think that all the auto-filled results that aren't different by design should be checked, not only the user provided results. The intent would be to test that all logged results (including auto-filled) are invariant to the which API is used, i.e., class API vs function API, excluding fields that are meant to be unique(-ish) to each trial or result, e.g., PID, experiment_id, timestamp.

Things like timesteps_*, episodes_*, or training_iteration, should probably be tested for consistency with regards to when/how they are added and updated across APIs even when not provided by the user. It might be worth re-opening this issue until the checking method is updated. What do you think?

[richardliaw]

Yeah that's actually a good point; it should actually be quite an easy fix. Let's reopen this issue for now.

[richardliaw]

Fixed in #4421.