Monotonicity of bisectors wrt precision

[gchabert]

In the current definition of a bisector, a "precision" parameter is given and it is said that when the domain of a variable is less than its corresponding precision, the bisector skips this variable

This has the unpleasant consequence that by skipping variables "under the precision", we lose monotonicity of a search tree with respect to precision. That is, the search tree obtained for a precision eps has no reason to be a subset of the search tree obtain for eps'<eps.

And, in practice, when one tries to run ibexsolve with a very coarse precision, the behaviour may be really bad because the bisector may skip important variables and keep on bisecting useless ones. In a word, a crude precision makes the bisector loses its "reason for being" as it only selects large-domain variables despite of its own heuristic.

As example, it was observed in #398 that the search tree for a crude precision (like 0.01) is much larger than for a thin one (like 1e-8).

Fix: Shall we remove this rule? That is: no precision anymore associated to a bisector. The only way it throws NoBisectableVariable is when it reaches 1 ulp. By the way, the precision of a bisector was redundant with that of the solver.