display an estimation of the best objective that can be expected from search

[sonichi]

Discussed in https://github.com/microsoft/FLAML/discussions/311

^{Originally posted by jezerro November 24, 2021} Hello, im working on a gesture recognition app that uses tabular data. As time passes im adding more gestures and its hard having to trial-and-error the balance between training time and samples per label.

It would be very helpful if the system decided how much time to run automatically and show a message to increase samples when they are not enough

Applicable to both AutoML.fit() and tune.run().