The Mamba in the Llama: Distilling and Accelerating Hybrid Models