refine benchmard bert ips stat

[wanghuancoder]

PR types

Bug fixes

PR changes

Others

Description

Benchmark统计平均每个step的IPS存在错误：
由于Bert模型在loss.numpy()时才会阻塞CPU，等待GPU完成全部运算。原有时间统计逻辑会遗漏从train_run_cost = time.time() - batch_start到logger.info（print loss）之间的GPU用时。
因此，相应的step只统计了前向的batch_cost，没有统计反向和Optimizer。

这种情况，如果log_step==1，那么没个step统计都是错误的。如果log_step==10，则每10个step中第1个step是错误的，后9个是正确的。这导致我们的IPS统计偏高。

但Torch的统计存在的问题，考虑到IPS更多用于与竞品性能对比使用。因此对该问题不做修改。但添加Note标注。

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 55.36%. Comparing base (2273ee7) to head (e083057).
Report is 58 commits behind head on develop.

❗ Current head e083057 differs from pull request most recent head 6962b99. Consider uploading reports for the commit 6962b99 to get more accurate results

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #8361      +/-   ##
===========================================
+ Coverage    55.15%   55.36%   +0.21%     
===========================================
  Files          601      614      +13     
  Lines        91764    96016    +4252     
===========================================
+ Hits         50611    53164    +2553     
- Misses       41153    42852    +1699     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[wawltor]

LGTM