Support Deepseek w4a8 quantization

[pichangping]

What this PR does / why we need it?

Does this PR introduce any user-facing change?

How was this patch tested?

[wangxiyuan]

Thanks for your contribution. I have some question before deep review:

1. how to convert a w4a8 weight? which version of modelslim should be used?
2. does deepseek v2 lite works with w4a8 as well? If yes, the e2e test should be added as well. @22dimensions please help to update a weight to modelscope as well.
3. does w4a8 only works with deepseek, how about any other model, like Qwen?

[pichangping]

1. First, quantize using the BF16 format to int8, then quantize the obtained int8 to int4. Since there is no int4 data format, the tool outputs the quantized weights using the low-precision int8.
   2.Whether it is compatible is not clear
   3.Other model features are supported, but accuracy is not guaranteed.

[pichangping]

In the main branch of this repository "https://gitee.com/ascend/msit", you can access the file "msmodelslim/example/DeepSeek/quant_deepseek_w4a8.py". Use the command "python quant_deepseek_w4a8.py --model_path DeepSeek-R1-BF16 --save_path DeepSeek-R1-W4A8 --mindie_format". If you are unfamiliar with the parameters, you can read msmodelslim/example/DeepSeek/README.md.

[pichangping]

modelslim
Use branch master
Command: Refer to README.md for 运行前必检 and DeepSeek-R1 w4a8 混合量化
Reference command: python3 quant_deepseek_w4a8.py --model_path {float weight path} --save_path {W4A8 quantized weight path} --mindie_format
Since mindie_format generates a mindie format, some adaptation modifications are needed for vllm usage:
Rename quant_model_description_w8a8_dynamic.json to quant_model_description.json, and add the configuration "group_size": 256 in this file
Modify config.json: Change model_type to deepseek_v3; delete quantization_config

[wangxiyuan]

rebase to main please, and move this file to tests/e2e/multicard

[Yikun]

The file name was also wrong....

[Yikun]

1. Please do rebase
2. Please update commits msg to show step by step test include

After this PR merged

1. Please do refactor to inhert w8a8
2. Please add doc

[Yikun]

Please remove this, this already remove in #1229