fixed bug in the "fast_descrs" optimization;
the bug is that two sets may actually be the same but because of the three-valued logic a different "The" or "Eps" is chosen; e.g. consider the set {1, 2}. If it is approximated in one place as {1, 2?} and in another place as {1?, 2}, then "Eps" would return 1 in the first case and 2 in the second case. This is of course wrong, because both sets potentially represent {1, 2}.
The current fix has a very negative impact on precision.