Several minor tweak to vllm benchmark workflow

[huydhn]

Several small tweaks:

• Disable arm benchmark for now as it's not working and there is no owner
• Only need to parse the list of model from serving benchmark config. It always has all the models because serving benchmark is the most basic one, and it has tensor_parallel_size field set correctly. This is to fix a non deterministic issue where a model's tensor_parallel_size could be missing when glob decides to return the list of benchmark configs differently
• Run daily, saving the capacity for the upcoming nightly run