Compiling LLMs into a MegaKernel: A path to low-latency inference