Optimize SmoothQuant alpha per-topology for the best accuracy on Llama2 family models

[vvchernov]

• Use different datasets for calibration (dummy, Pile, gsm8k, triviaqa ans so on)
• Use llama2-7b with different int8 quantization types
• Use alpha in range (0, 1)
• Use lm-evaluation-harness to accuracy benchmark on tasks like gsm8k, triviaqa ans so on

[Ailurus1]

Current results for Llama2-7b accuracy testing with different $\alpha$ parameter are here