Policy engine dependency graph is slow

Every insert or delete into any policy results
in re-computation of that policy's dependency graph and
a check that no recursion occurred either within a policy
or across policies. This dependency graph computation was
expensive because (i) instead of updating the graph based
on the policy statements inserted/deleted, the graph
was recomputed from scratch and (ii) each policy had
its own dependency graph, and to check for recursion
across policies we computed the union of those individual
graphs. In addition, we were copying the list of current
rules before doing any of that computation.