DeepSeek; modify scale save and quantization config; fix thinking data calibration

[ali-88123]

• 修改了scale的保存方式：同层scale和weight存储在同一个safetensors中
• DeepSeek保存类重命名：单卡运行保存方式为DeepSeekV3PTQSaveSingle，多卡为DeepSeekV3PTQSaveMulti

def get_save_func(self):
    if self.deploy_backend in ["vllm", "trtllm"]:
        if self.model.using_multi_nodes:
            return DeepSeekV3PTQSaveMulti
        return DeepSeekV3PTQSaveSingle
    else:
        raise NotImplementedError(
            f"deploy_backend {self.deploy_backend} is not supported for saving."
        )

• 更新了quantization_config
• 修复了TextDataset无法处理conten包含数据的问题
• 删除merge_model以及tp_model暂存文件夹