BiTA: Bi-Directional Tuning for Lossless Acceleration in Large Language Models