Replace Relu with bounded Relu in MobileNetV2 quantization mkldnn

[wozna]

It removes the workaround from PR #17570

In MKL-DNN 0.18 there was no support for bounded ReLU post-op. Hence, when adding support for MobileNetV2 support we decided to use ReLU instead of ReLU6 (bounded relu).
As the MKL-DNN version is upgraded to 0.20 now and it supports bounded ReLU post-op, the workaround can be removed and bounded ReLU can be used.

test=develop

[wozna]

@Sand3r- I am asking for a review

[bingyanghuang]

@lidanqing-intel could you help review this PR?

[lidanqing-vv]

Looks good to me~
Later we may consider to call DequantizeOutput after conv_op->Op()->SetAttr("fuse_brelu_threshold", scale_out * threshold) , not before, because DequantizeOutput uses conv_op. But anyway conv_op is passed by pointer so it works well.

[luotao1]

LGTM