Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels