Narrowing from widening thresholds

[Red-Panda64]

This PR allows narrowing of integer intervals when the bound being narrowed is a widening threshold. Beforehand, when an interval was widened to a threshold and the threshold fit (e.g. widen [0, 0] [0,1] = [0, 10] for an upper threshold 10), it could no longer be narrowed (even if []). This is because int intervals implement accelerated narrowing and will only narrow from from maximal upper or minimal lower bounds.
This PR will use the more precise bound, only when old bound is maximal/minimal or an upper/lower threshold. The narrowing must still terminate, since upper/lower bounds can only shrink/grow as often as there are thresholds.

Example:

int main() {
    int i;
    for(i = 0; i < 10 && i < 20; i += 3);
    __goblint_check(i <= 12);
}

The i < 20 condition is only in this example to be identified as an upper threshold.
Without this PR, the invariant cannot be shown when wideningThresholds are enabled: i is first widened to [0, 10], then [0, 20] at the loop head. Then, the narrowing iteration cannot reduce i to the more precise interval [0, 12]. With these changes, the more precise interval is inferred.

[sim642]

Interval sets also use the widening thresholds, so it would make sense for them to also support this logic, if they already don't.

[Red-Panda64]

Good point. I've made that adjustment. Also, I have fixed an oversight I made: before, the narrow operator did not have to explicitly check whether the new bound is more precise, because it would only change the bound if the previous value was maximal/minimal. Now it does have to check.

[Red-Panda64]

I've also now added a unit test to make the modified narrow operator does indeed narrow from thresholds and does not permit growth. The test applies to intervals and interval sets, as well as upper and lower bounds.