ENH limit the way we grow the stop_val

We have a very aggressive way to increase the `stop_val`, which can lead to very long run, blowing up the timeout for no good reason.
The way we get the next `stop_val` should take into account the time to compute the objective, the time left to the timeout and the time taken to perform the last iteration and find a good balance between them.

This would definetly be a plus, and avoid the timeout blow up that are a pain when running the benchmark (typically, timeout set to 50min but the run takes more than 75min because the too many iteration are run).