Kernel Machines Beat Deep Neural Networks on Mask-based Single-channel Speech Enhancement