Towards Modular LLMs by Building and Reusing a Library of LoRAs