Understanding Optimal Feature Transfer via a Fine-Grained Bias-Variance Analysis