[tune] Support extending Async Hyperband runs past max_t

[ANarayan]

Would it please be possible to extend the async hyperband scheduler implementation to have the best n trials train to completion beyond max_t and stop based on some other stopping criteria (i.e. TrialPlateauStopper)? Currently, the scheduler stops training a given trial when it is the last one left in its bracket. Hence, it wont actually return a fully trained model that has minimized or maximized the metric in question.

A similar feature was implemented for the BOHB/Hyperband scheduler here.

Having access to the fully trained models and knowledge of their final performance metrics would be very useful for benchmarking and comparative analysis tasks.