Optimize/cleanup array graph domain

[caballa]

From @caballa on April 9, 2017 8:34

The current implementation works for toy programs but two main things
must to be done for being able to analyze real programs:

1. landmarks must be kept as local state as part of each abstract
   state.

2. reduction between scalar and weight domains must be done
   incrementally. For that, we need some assumptions about the
   underlying scalar domain. For instance, if we assume zones then
   after each operation we know which are the indexes affected by
   the operation. We can use that information for doing reduction
   only on those indexes. This would remove the need of having
   methods such as array_sgraph_domain_traits::is_unsat and
   array_sgraph_domain_traits::active_variables which are anyway
   domain dependent.

Copied from original issue: caballa/crab#15

[caballa]

Resolved in a way by commit 5e3fac4