Benchmarking update - phase 1

[dbarbuzzi]

This PR updates the benchmarking performed in remote-push and nightly runs according to the first set of deliverables from our recent meeting:

• Only the benchmark_serving.json config is run
  • This is accomplished with a new list, nm_benchmark_base_config_list.txt, other lists are untouched
• The benchmark_serving.json has various reductions:
  • Model list reduced to facebook/opt-350m and meta-llama/Meta-Llama-3-8B-Instruct
  • nr-qps list reduced to 300,1
  • Metric tracking reduced to mean TPOT and mean TTFT (other metrics still recorded/logged per usual)

There is also a small fix related to server startup (changing from localhost to 127.0.0.1 because localhost on the machines is mapped to the IPv6 ::1 which something in the server stack doesn’t seem to like).

In a commit prior to opening the PR with all functional changes, the full benchmark job took <30 min:
https://github.com/neuralmagic/nm-vllm/actions/runs/9669361155/job/26709082658

[andy-neuma]

thanks