Toward Efficient Low-Precision Training: Data Format Optimization and Hysteresis Quantization