Add conv dequant squash

[wozna]

This squash is based on PR #18754
This squash change chain with conv op. and dequantize op. to conv with forced float output.
Thanks to this you will avoid additional conversion to int8.

In mobilenet ssd there are 12 pattern like this and the use of this squash improves accuracy
(on clx machine):
- Before squash: acc_fp32 = 0.7733, acc_int8=0.77
- After that squash: acc_fp32 = 0.7733, acc_int8=0.7719

Unfortunately, after a few tests using mkl-dnn and int8 on clx machine, fps decreased:
- Before squash: avg fps = 151.634
- After squash: avg fps = 146.8749
Is's almost 3%.
That could be some kind of option if accuracy is more important than performance.

[wozna]

@luotao1 what do you think about the result and option to enable the pass?

[bingyanghuang]

Tested on Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHz

• Test command line
  ./paddle/fluid/inference/tests/api/test_analyzer_int8_object_detection "ARGS" --infer_model=third_party/inference_demo/int8v2/mobilenet-ssd/model --infer_data=/home/bingyang/.cache/paddle/dataset/pascalvoc/pascalvoc_full.bin --warmup_batch_size=10 --batch_size=1 --paddle_num_threads=1

• Before

I0821 06:37:47.878208 109578 tester_helper.h:468] --- Performance summary ---
I0821 06:37:47.878234 109578 tester_helper.h:469] FP32: avg fps: 15.5156, avg latency: 64.4513 ms
I0821 06:37:47.878245 109578 tester_helper.h:472] INT8: avg fps: 20.4021, avg latency: 49.0146 ms
I0821 06:37:47.880376 109578 tester_helper.h:452] --- Accuracy summary ---
I0821 06:37:47.880390 109578 tester_helper.h:453] Accepted mAP drop threshold: 0.01. (condition: (FP32_mAP  - INT8_mAP ) <= threshold)
I0821 06:37:47.880398 109578 tester_helper.h:457] FP32: avg mAP 0.7392
I0821 06:37:47.880403 109578 tester_helper.h:459] INT8: avg mAP 0.7287

• After

I0821 05:35:45.883280 75832 tester_helper.h:468] --- Performance summary ---
I0821 05:35:45.883306 75832 tester_helper.h:469] FP32: avg fps: 15.4869, avg latency: 64.5709 ms
I0821 05:35:45.883317 75832 tester_helper.h:472] INT8: avg fps: 21.2940, avg latency: 46.9617 ms
I0821 05:35:45.885444 75832 tester_helper.h:452] --- Accuracy summary ---
I0821 05:35:45.885458 75832 tester_helper.h:453] Accepted mAP drop threshold: 0.01. (condition: (FP32_mAP  - INT8_mAP ) <= threshold)
I0821 05:35:45.885465 75832 tester_helper.h:457] FP32: avg mAP 0.7392
I0821 05:35:45.885469 75832 tester_helper.h:459] INT8: avg mAP 0.7315

• Benchmark conclusion

Compare	FPS INT8	Accuracy
Before	20.4021	0.7287
After	21.2940	0.7315

[luotao1]

@juncaipeng @wzzju Please double-check the #18905 (comment)

[wojtuss]

LGTM

[wozna]

@luotao1 if everything is fine then this PR is an improvement ready to merge

[juncaipeng]

Tested on Intel(R) Xeon(R) Gold 6271 CPU @ 2.60GHz

• Test command

./test_analyzer_int8_object_detection "ARGS" --infer_model=mobilenet-ssd/model/ --infer_data=/dev/shm/pascalvoc_full.bin --warmup_batch_size=10 --batch_size=1 --paddle_num_threads=1

• Before

I0827 13:02:15.814708 370741 tester_helper.h:471] --- Performance summary ---
I0827 13:02:15.814764 370741 tester_helper.h:472] FP32: avg fps: 12.3974, avg latency: 80.6618 ms
I0827 13:02:15.814790 370741 tester_helper.h:475] INT8: avg fps: 15.9749, avg latency: 62.5980 ms
I0827 13:02:15.817037 370741 tester_helper.h:455] --- Accuracy summary ---
I0827 13:02:15.817065 370741 tester_helper.h:456] Accepted mAP drop threshold: 0.01. (condition: (FP32_mAP  - INT8_mAP ) <= threshold)
I0827 13:02:15.817085 370741 tester_helper.h:460] FP32: avg mAP 0.7392
I0827 13:02:15.817091 370741 tester_helper.h:462] INT8: avg mAP 0.7287

• After

I0827 11:10:15.486403 161780 tester_helper.h:468] --- Performance summary ---
I0827 11:10:15.486446 161780 tester_helper.h:469] FP32: avg fps: 12.4536, avg latency: 80.2981 ms
I0827 11:10:15.486469 161780 tester_helper.h:472] INT8: avg fps: 16.6916, avg latency: 59.9104 ms
I0827 11:10:15.488651 161780 tester_helper.h:452] --- Accuracy summary ---
I0827 11:10:15.488679 161780 tester_helper.h:453] Accepted mAP drop threshold: 0.01. (condition: (FP32_mAP  - INT8_mAP ) <= threshold)
I0827 11:10:15.488698 161780 tester_helper.h:457] FP32: avg mAP 0.7392
I0827 11:10:15.488703 161780 tester_helper.h:459] INT8: avg mAP 0.7315

• Benchmark conclusion

Compare	FPS INT8	Accuracy
Before	15.9749	0.7287
After	16.6916	0.7315