On the Connection between $L_p$ and Risk Consistency and its Implications on Regularized Kernel Methods