Feature Alignment: Rethinking Efficient Active Learning via Proxy in the Context of Pre-trained Models