Retro: Reusing teacher projection head for efficient embedding distillation on Lightweight Models via Self-supervised Learning