feat: add w8a8_dynamic quant & support deepseek quant

[zzzzwwjj]

add w8a8_dynamic quant & support deepseek quant

[MengqingCao]

Try to record the reason why we patching deepseekv2 model, please add if there are any omissions:

1. load offset weight in fused_moe_perchannel_weight_loader, the original weight loader in vllm 0.7.3 could only load scale and zero
2. flatten weight to adapt npu ops

            if "scale" in name or "offset" in name:
                loaded_weight = loaded_weight.flatten()

3. add packed_modules_mapping in deepseekv2 model so that the gate proj and up proj could be splitted

TODO:

• Before we merging this to branch main, we need to remove the fused_moe_perchannel_weight_loader, as
  [moe][quant] add weight name case for offset vllm#15515 has been merged

[MengqingCao]

LGTM, thx!