Change cost_attr when max_iter is used as the stopping criterion instead of time_budget in AutoML.fit

[sonichi]

Discussed in #779

^{Originally posted by flippercy October 28, 2022}
Hi Team FLAML:

I have a few general questions regarding FLAML:

1. Could I use a customized learner for the ensemble? For example, for a linear combination, I want to make sure all the coefficients are negative.

2. I noticed that the search of automl is much more biased when running distributedly on compute clusters, which means that it focuses much more on 'strong' learners. For example, this is the summary from my recent search:
   
   Obviously FLAML spent the majority of time on LearnerE, which is the optimal one; it is understandable, however, the distribution of attempts on different learners is much more balanced when running FLAML on a VM. Could you explain this phenomenon, please? Usually we use FLAML not just to look for the overall optimal solution, but also the best result per learner.

Thank you!

If we set cost_attr for BlendSearch as None, then the number of iterations will be used as cost measurement. It's supposed to allocate # trials to different learners in a more balanced manner in the parallel setting.

[sonichi]

closed by #837