[CPU] Optimize MaxPool2D and AvgPool2D.

[mciprian13]

Summary

• Optimize MaxPool2D and AvgPool2D (around 30 to 40 % faster for large number of channels) by removing the use of libjit_getXYZ sort of macros for accessing tensors.
• Unify the AvgPool2D CPU quantized kernels for both flavors of the CountIncludePads and use only fix-point arithmetic (previously for the CountIncludePads false some floating-point was used)
• Treat the special case when the padding is larger than the pooling window size (when the pooling window includes only padding pixel then we return 0 for that window).

Test Plan
Add unit tests for the special case when the padding is larger than the pooling window size.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 15 days if no further activity occurs. Thank you for your contributions.

[mciprian13]

ping!

[facebook-github-bot]

@842974287 has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[842974287]

LG! Do you mind also add these operator tests in the NNPIOperaterTest blacklist as well.

[mciprian13]

@842974287 I also added the unit tests to NNPI black list. Can we have this landed then? Also could you please take a look on #4971. Thanks a lot for your time!

[facebook-github-bot]

@842974287 has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[842974287]

@mciprian13 No problem, I'll take a look on #4971. Thanks a lot for doing this!

[facebook-github-bot]

@842974287 merged this pull request in 6bd0b20.