Documentation about small budget

Thank you for your explaination @sonichi !
I now understand it. For small time budgets, It is advisable to rather use a predifined `max_iter` than a `time_budget`.
@qingyun-wu : That is an interesting design question. I think returning a `NoneType` might break the code for some users. However, it is also a very special case that someone uses that little training time. Maybe a bit more explaination about that in the documentation might be helpful. For me a dummy regressor / classifier would make sense too, though. :)

_Originally posted by @OliverSchacht in https://github.com/microsoft/FLAML/discussions/738#discussioncomment-3794757_