Scaling Sentence Embeddings with Large Language Models