Improve cached_op performance for static mode

[ZhennanQin]

Description

@pengzhao-intel @TaoLv @xinyu-intel @junrushao1994
When gluon model hybridize with static_shape=True, static_alloc=True, cached_op with static mode will be used. For this situation, we should try to cache operator state for better performance. This PR is to enable this feature to speed up gluon inference speed, especially for small batch sizes.

Below data is collected on SKX-8180 28 cores, SKX GLUON INT8 OPT shows the performance change from this PR, base is SKX GLUON INT8.

GLUON ResNet50 V1(10 cores for rec decoder)	SKX GLUON FP32	SKX GLUON INT8	SKX GLUON INT8 OPT
Throughput(img/sec, bs=1)	64.6	11.89	144.53
Throughput(img/sec, bs=2)	79.74	23.1	226.04
Throughput(img/sec, bs=4)	111.9	43.45	302.4
Throughput(img/sec, bs=8)	134.86	78.72	347.6
Throughput(img/sec, bs=16)	143.52	129.09	362.93
Throughput(img/sec, bs=32)	146.63	197.63	381.2
Throughput(img/sec, bs=64)	153.89	261.33	380.79
Throughput(img/sec, bs=128)	156.82	326.23	408.38
Accuracy(5000 imgs)	77.21%93.55%	76.86%/93.46	76.86%/93.46

GLUON MobileNet1.0(10 cores for rec decoder)	SKX GLUON FP32	SKX GLUON INT8	SKX GLUON INT8 OPT
Throughput(img/sec, bs=1)	166.23	38.77	281.28
Throughput(img/sec, bs=2)	238.79	75.81	518.52
Throughput(img/sec, bs=4)	333.2	143.83	987.63
Throughput(img/sec, bs=8)	397.46	262.47	1245.85
Throughput(img/sec, bs=16)	425.35	425.63	1332.25
Throughput(img/sec, bs=32)	451.89	653.8	1474.7
Throughput(img/sec, bs=64)	471.77	897.99	1528.63
Throughput(img/sec, bs=128)	468.67	1125.75	1557.16
Accuracy(5000 imgs)	73.28%/91.22%	72.85%/90.99%	72.85%/90.99%

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http:https://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Feature1, tests, (and when applicable, API doc)
• Feature2, tests, (and when applicable, API doc)

Comments

• If this change is a backward incompatible change, why must this change be made.
• Interesting edge cases to note here

[pengzhao-intel]

cc @zhreshold

[zhreshold]

Please fix CI and minor issue, this is awesome!!

[larroy]

Can the performance measurement scripts be shared?

[pengzhao-intel]

@larroy I think the test is already part of gluonCV
https://gluon-cv.mxnet.io/build/examples_deployment/int8_inference.html
@xinyu-intel am I right?

[xinyu-intel]

@pengzhao-intel @larroy yes, mainly use imagenet classification verify_pretrained.py and eval ssd for gluoncv evaluation. BTW, after Gluoncv #755 merged along with this pr. The performance will get improvement.

[pengzhao-intel]

@zhreshold please help to review again :)

[pengzhao-intel]

LGTM

[zhreshold]

lgtm