Lobsters: Native-speed vLLM transformers modeling backend