APRICOT: Active Preference Learning and Constraint-Aware Task Planning with LLMs