Attend Before Attention: Efficient and Scalable Video Understanding via Autoregressive Gazing