State tracking of constraints.

[rkaminsk]

None of the constraints in clingcon can really benefit from the calculation of lower/upper bounds during propagation (even though it helped to find a lot of bugs). Even for sum constraints propagation is still linear (or quadratic if counting conflict generation). So we can simplify this by just enqueuing a constraint. This would also allow for unwatching a constraint on the current decision level and not just the next one.