Refactor PerfZero for simplicity and ease-of-use

[lindong28]

This patch makes the following improvement. It depends on cl/230280956 which adds leading_indicators_test.py to OSS tensorflow/benchmarks.

Add leading_indicators_test.py to benchmarks/scripts and run it in PerfZero
Read benchmark results from local protobuf file written by tf.test.Benchmark.report_benchmark()
Simplify Perfzero report logic implementation
Rename environment variables from ROGUE_* to PERFZERO_*
Replace print(..) with logging.info(..) and logging.debug(..)
Print benchmark summary in a human readable format
Use datetime string as execution id and include it in the output path name
Specify full list of environment variables with documentation in README.md
Configure bigquery table name via environment variable
Upload gpu driver version to bigquery
Print messages to both stdout and log file

benchmark summary is {
  "system_info": {
    "platform_name": "workstation-z420",
    "cpu_socket_count": 1,
    "gpu_count": 2,
    "cpu_model": "Intel(R) Xeon(R) CPU @ 2.20GHz",
    "cpu_core_count": 8,
    "gpu_model": "Tesla V100-SXM2-16GB"
  },
  "execution_id": "2019-01-29-04-25-52-683031",
  "benchmark_info": {
    "output_url": "gs://tf-performance/test-results/2019-01-29-04-25-52-683031/",
    "date_time": "2019-01-29-04-26-45",
    "project_name": "perfzero-dev"
  },
  "benchmark_result": {
    "success": false,
    "benchmark_name": "Resnet50Benchmarks.benchmark_fake_1gpu_gpuparams",
    "wall_time_ms": 49606,
    "iters": 100,
    "loss": "5.864147",
    "top_5_accuracy": null,
    "top_1_accuracy": null
  },
  "ml_framework_info": {
    "framework": "tensorflow",
    "version": "1.13.0-dev20190128",
    "git_version": "v1.12.0-6820-g0f59cbd297",
    "build_type": "OTB",
    "channel": "NIGHTLY"
  }
}

The information above is translated to old format and uploaded to bigquery.

The PerfZero configuration is now documented in README.md

[tfboyd]

Stream should be the default and needs to exist. I think you are swapping back in the current reporting. I think it makes sense to not change the reporting structure until we have a strong need. Changing it does not get us a lot of value right now and adds a lot of work. I realize the current structure is a bit messy right now.

[tfboyd]

I left this off as I was testing pytorch and mxnet with the same setup and wanted to share reporting. I could be talked into keeping this here but I kind of prefer we not require tensorflow to do reporting.

[tfboyd]

If you do keep this you need a pylint disable I think.