Adaptive Principal Components Allocation with the $\ell_{2,g}$-regularized Gaussian Graphical Model for Efficient Fine-Tuning Large Models