[V1 Loader] Support qwen2(bf16)

[zeroRains]

pcard-71500

注意：V1 Loader用到了Paddle develop版本对set_value和copy_的新特性，因此需要使用Paddle develop或者3.2(暂未发布)及以上的版本。

本PR为Qwen2模型适配V1 Loader加载。
目前在bf16下验证，精度与develop对齐，且保证旧版和新版的load都能正常工作。
接入V1 Loader后，在Qwen2-7B-Instruct中，TP4的模型加载时内存占用下降为原本的29%，加载时间持平。

修改内容：

1. 在Qwen2模型中新增load_weights方法，以适配V1 loader。
2. 在QKVParallelLinear中新增bias_loader方法，以便在bias调用weight_loader时，也有对应的方法执行。

测试记录：
模型：Qwen2-7B-Instruct
性能记录

指标	旧版loader	V1 Loader
TP4内存占用（GB）	17	5
TP4模型加载时间（s）	4.41	4.41

ps: 内存占用指从fd开始执行端到端推理到结束出现的内存峰值。模型加载时间是指各个rank的worker.log中记录的模型加载时间的最大值。

精度记录

Qwen2-7B-Instruct tp1 bf16 结果对齐
base bf16 : ? I am a large language model created by Alibaba Cloud. I am called Qwen.\n\nCan you generate a poem about the beauty of nature?
pr   bf16 : ? I am a large language model created by Alibaba Cloud. I am called Qwen.\n\nCan you generate a poem about the beauty of nature?

tp4 bf16 结果对齐
base bf16 : ? I am a large language model created by Alibaba Cloud. I am called Qwen.\n\nCan you generate a poem about the beauty of nature?
pr   bf16 : ? I am a large language model created by Alibaba Cloud. I am called Qwen.\n\nCan you generate a poem about the beauty of nature?

测试脚本：

# run_demo.sh

export CUDA_VISIBLE_DEVICES=0,1,2,3
export PYTHONPATH=/workspace/miniconda3/envs/py310:$PYTHONPATH
export PYTHONPATH=/workspace/FastDeploy:$PYTHONPATH
python offline_demo.py

# offline_demo.py
from fastdeploy.engine.sampling_params import SamplingParams
from fastdeploy.entrypoints.llm import LLM

model_name_or_path = "/workspace/Models/Qwen2-7B-Instruct"
sampling_params = SamplingParams(temperature=0.1, max_tokens=30, top_p=0)
# load_choices取值"default_v1"是新版Loader，"default"是旧版
llm = LLM(model=model_name_or_path,num_gpu_blocks_override=1024, tensor_parallel_size=4, load_choices="default_v1")
output = llm.generate(prompts="who are you",
                      use_tqdm=True,
                      sampling_params=sampling_params)
print(output)

[paddle-bot]

Thanks for your contribution!

[yuanlehome]

加一个CI单测吧

[zeroRains]

加一个CI单测吧

这个跑端到端模型推理会好处理一些，直接将这部分抽离出来写单测，不太好弄。