Drastically improve performances by simplifying the book-keeping

Drastically improve performances by simplifying the book-keeping
of pending transitions during a topdown run.
Before:
      - we were keeping for each node
        (1) sat: the set of states satisfied at that node
        (2) unsat: the set of states unsatisfied at that node
        (3) todo: the hashconsed list of pending transitions (of the
            form q, {a,b,c} -> f
            where f has been simplified w.r.t to sat and unsat (that is
            atoms containing states in either have been simplified to True
            or False) and q is neither in sat nor unsat.
      - during each run, for each node, we were scanning todo and rebuilding
        it removing satisfied or unsatisfied transitions. This lead to excessive
        hashconsing, and a lot of time spent in the hashconsing module (in turn
        spending a lot of time in the GC)
After:
       - we keep for each node:
         (1) sat: the set of states satisfied at that node
         (2) todo: the set of undetermined *states*
       - we perform the book-keeping as before but do not materialize the
         simplified transitions. Also unsat is made inplicit (a state is
         unsat if it is neither in sat nor todo).