Add kvint4/8 ete testcase

[zhulinJulia24]

1. benchmark.yaml change

• add kvint4/8 benchmark
• change cuda allocate for more effectiveness

2. daily_ete_test.yaml change

• remove kvint8 quantization
• remove timeout setting and reorder teststep because the testcase is much more stable

3. evaluate.yaml change

• change model path

4. pr_test_case chage

• remove -s -v to make result more simpler and clearer

5. autotest/chat_prompt_case.yaml and autotest/prompt_case.yaml

• react testcase to reduce error results

6. autotest/config.yaml and autotest/utils/config_utils.py

• split models into chat model and base model and add more models
• refactor kvint setting

7. autotest/interface/restful/test_restful_interface_func_common.py

• add more testcase into restful interface

8. autotest/tools/chat/*

• Streamline test cases, use one test case to cover multiple previous test cases
• split testcases into chat and base model

9. autotest/tools/pipeline/*

• add kvint4/8 testcases

10. autotest/tools/restful/*

• add kvint4/8 testcases
• refactor start and stop restful server process

other changes:

• rename -inner-w4a16 models into -inner-4bits
• change temperater=0.01 into top_k=1 to make response more stable
• remove useless code after refactor kvint4/8 quantization

[zhulinJulia24]

test result

daily ete testcase:

https://github.com/zhulinJulia24/lmdeploy/actions/runs/8720843216
some response in some model is still not stable. I will still follow and improve it.

benchmark testcase:

https://github.com/zhulinJulia24/lmdeploy/actions/runs/8724161035

evaluation testcase:

https://github.com/zhulinJulia24/lmdeploy/actions/runs/8720159669

and
https://github.com/zhulinJulia24/lmdeploy/actions/runs/8724058314

[zhulinJulia24]

pr_test time cost reduce to about 30mins, because of streamline test cases, use one test case to cover multiple previous test cases in cli chat. Currently 4bits quantization is still reserved.

[lvhan028]

pr_test time cost reduce to about 30mins, because of streamline test cases, use one test case to cover multiple previous test cases in cli chat. Currently 4bits quantization is still reserved.

How long did pr_test workflow cost before this PR?

[zhulinJulia24]

pr_test time cost reduce to about 30mins, because of streamline test cases, use one test case to cover multiple previous test cases in cli chat. Currently 4bits quantization is still reserved.

How long did pr_test workflow cost before this PR?

currently about 1hour.

[zhulinJulia24]

--no-deps fixed

[zhulinJulia24]

benchmark step condition change test ok:

[RunningLeon]

LGTM