Meta-Learning Action Conventions In Ad-Hoc Hanabi