Add warmup time and compile time log for the eval/prediction.

[jiminha]

What does this PR do?

We are adding log prints for compile time in our evaluation/prediction loop for all inference test cases. To get the compile time, use the --throughput_warmup_steps flag (same logic as in text_generation/run_generation.py). This will also increase the throughput numbers since it will remove the warmup time from the throughput calculation.

• examples/language-modeling/run_mlm.py

  ***** eval metrics *****
  ..
  eval_graph_compliation_duration =      0.949
  ..

• examples/summarization/run_summarization.py

  ***** predict metrics *****
  predict_gen_len                    =     5.4174
  predict_graph_compliation_duration =     9.3589
  ...

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.