Compress then Serve: Serving Thousands of LoRA Adapters with Little Overhead