Scaling Up Your Kernels: Large Kernel Design in ConvNets towards Universal Representations