QATv2 updates, minor bug fixes

[oguzhanbsolak]

Change log:
1- QATv2
2- ActionTCN, AutoEncoder, FaceID_112 and MobileFaceNet_112 models are refactored for symbolic_trace() compatibility
3- EfficientNetv2 performance optimizations
4- Bug fixes:
a- Imagenet.py -> transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) was causing most of the inputs to be out of range for ai8x.normalize()
b- Train.py -> m.are_locations_wide should be a function call
c- Train.py -> For 8 bits classification, wide tests, the output should be divided with 128 instead of 256. Also, for the wide check, named_modules() is more robust than dict search.
d- ai8x.py -> fuse_bn_layers() is using the named_modules() now, as it works with the models that have nn.sequential modules as well
e- ai8x-blocks.py -> Refactoring depthwise_conv for layer naming convention

[rotx-eva]

When updating README.md, please also update README.pdf and create a matching PR in ai8x-synthesis

[oguzhanbsolak]

Added "test with best checkpoint" functionality

[ermanok]

Looks good.