Learning Dense Correspondence for NeRF-Based Face Reenactment