[Perf](bench): refactor benchmark scripts for unified format

[PerryZhang01]

Motivation

This PR is for unified benchmark test. We provided unified run_server.sh and run_bench.sh scripts for deepseekv3.2, gptoss and llama to test benchmark, the results will be printed on the screen and saved in an excel file. Users only need to modified the configs such as MODEL_PATH in run_bench.sh.
screen:

excel :

Note: we only test deepseek3.2. Gptoss and llama are not test, anyone can modifiy their params in run_server.sh and run_bench.sh after testing.

[valarLip]

a little bit toooo much if we have scripts for each model..

[PerryZhang01]

every model has different config on server and client, scripts or recipes are necessary for them. users can directly reproduce the public performance using the scripts without worrying about special parameters and environment variables settings. everyone using the same scripts can reduce the performance gap, and can also used for CI.

[PerryZhang01]

removed gpt-oss and llama scripts, and only retained deepseek.