Length-Controlled Margin-Based Preference Optimization without Reference Model