Perf: analyze and reach parity with MLPerf LoadGen, vLLM benchmark_serving (and backend metric server)

[nvzhihanj]

We need to make sure what we measure is reflecting the real endpoint perf.

We need to cross compare inference endpoint reported perf/latency with the following on the same endpoint:

• MLPerf LoadGen (use 5.1 submission harness)
• benchmark_serving.py used by SemiAnalysis
• (Tentative) AIPerf