Following the Autoregressive Nature of LLM Embeddings via Compression and Alignment