Precise learning curves and higher-order scaling limits for dot-product kernel regression*