Random Features Outperform Linear Models: Effect of Strong Input-Label Correlation in Spiked Covariance Data