Feat: define and implement the final report structure

[nvzhihanj]

We already have the rough performance report. We need to define a structured performance and accuracy report so it's easy to be aggregated and parsed by the WebUI in the future

[mosalov]

Right now the benchmark command generates two versions of reports:

1. the console output consisting of Summary and Latency Breakdowns,
2. the content of the file that can be passes into the --output parameter.

Looking to the content of both, I see that they even contain different numbers:

----------------- Summary -----------------
Total samples issued: 640
Total samples completed: 640
Duration: 102.73 seconds
QPS: 6.23

vs

  "results": {
    "total": 640,
    "successful": 640,
    "failed": 0,
    "elapsed_time": 104.19038772583008,
    "qps": 6.14260119353924
  }

so most likely they are generated by two unrelated pieces of the code.
I do not think it should be so.

[pgmpablo157321]

@nvzhihanj @mosalov
Ideally, from the side of the UI, we can have performance logs that are as similar as possible as the ones generated by LoadGen (at least for inference v6.0).
For the accuracy report, there is not a standard format, but it should contain the metric and be able to be read by the submission checker

[nvzhihanj]

@nv-alicheng who will address this