DiJiang: Efficient Large Language Models through Compact Kernelization