Evaluate vLLM / SGLang benchmarks for baseline when making EzDeploy config changes

[stikkireddy]

• use one of the benchmarks on either 1xA100 or 2xA100 and confirm improvements in benchmarks before the following:

  • Upgrading versions of vllm or sglang
  • Modifying ez-config changes for quantization, spec decode, prefix caching, or block size defaults

• tests should be simple notebooks that can be scheduled on a job running the base engines on a single node vm.

• the goal is not to benchmark e2e latency but identify perf issues when making config changes.