Fixed EfficientNetV2's MBConvBlock output phase.

[sebastian-sz]

Fixes keras-team/tf-keras#329

Unindents output phase of MBConvBlock so that it is applied regardless of SE ratio.

To my knowledge, when inspecting the EfficientNetV2's configs, all MBConvBlocks use SE block part so the models and weights should be unchanged.

[fchollet]

Thanks for the PR. What validation accuracy do you see with the modified model on the imagenet test set?

CC @ianstenbit

[sebastian-sz]

@fchollet I did not run Imagenet evaluation as I suspect the models are unchanged. For all variants, after fix:

1. The existing weights load without issue.
2. If running inference on the same image, these models output the same values as the ones in tf-nightly.

gist I used for comparison.

[ianstenbit]

Thanks Sebastian and Francois -- looks like this fix will also need to be made in KerasCV's EfficientNetV2 (and it will be in keras-team/keras-cv#1146)

[haifeng-jin]

Pending on another reply from @fchollet .
Hold till 01/09/2023 to add the pending label again if no response.

[fchollet]

LGTM, thanks!