Building complete abstract interpretations in a linear logic-based setting