Prior preferences in active inference agents: soft, hard, and goal shaping