[las] better analyze maxnorms
Imported issue: Initially reported by @zimmerma in https://gforge.inria.fr/tracker/?group_id=2065&aid=14990
currently we compute the maximal norm on each side, and scale correspondingly the (estimated) norms
to use most of the 256 values we have in the sieving array.
However maybe we overestimate the bound, or we could allow to lose some relations of maximal norm to get a better accuracy in the relations of smaller norm (which are more likely to give relations).
For example with the sieving example from bug #14988 (closed) in this tracker, we get 736 relations.
If we reduce the maximal norm by a factor of 2 (before taking the log2) we get 744 relations (8 more).
If we reduce by a factor of 4, we get 739 relations (still 3 more). By a factor of 8: 736 (idem).
By a factor of 16: 733 (only 3 less).
A more detailed analysis should be performed, with a larger sieving region, but it seems we can get a few extra relations this way.
Paul