LoLCATs: On Low-Rank Linearizing of Large Language Models