Make key rank estimation bin size tuning smarter.

[cassiersg]

rank_accuracy has the surprising behavior of being sometimes very expensive even when there are few candidates. This happens when many keys have the exact same probability, such that the requested accuracy cannot be reached, and the algorithm reduces bin size until it reaches the max number of bins.

We should:

• have a way to stop reducing bin size early, when it is clear that reducing it won't improve our rank bounds,
• have a default max_nb_bin=None that raises an exception if the desired accuracy cannot be reached within a default bound on the number of bin. That would make the cost more explicit to the user (who can still remove the exception the setting max_nb_bin, and the exception could print the bounds it got.