Not yet implemented
-------------------

o Why is _symindex_map a hash_map and not directly linked from
  each variable as a pointer?? Don't recall. Might be more efficient
  to do it as links.

o Setting pb-watch pointers when a pb-constraint is added which
  has some lits already set, resulting in no valid watch set

o When variables are set at levels earlier than current_dl, new
  pb-implications are currently set at max_dl (the max of the levels
  of all false lits in the pbcons). Some of these might actually
  be implied even earlier, somewhere between the level dl at which the
  variable is being set (to false, of course) and max_dl (> dl).

o Memory usage for maintaining symmetry info

o pb learning on conflicts

o pb learning on symconflicts


Ideas for better performance
----------------------------

o Some pb-constraints have large rhs, putting most lits in the
  watch set. This makes searching for new watch lits inefficient.
  Consider having an "unwatched lits" vector/set for such constraints.
  (e.g. hole constraints in php - at least n-1 out of n pigeons don't go
  into any given hole.) Will have to maintain the set when unsetting
  vars, however. Not good.

