[BLOCKED] Replace IntermediateLayerGetter with feature_extraction.create_feature_extractor()

🚀 The feature

A new feature extraction utility based on FX was introduced by #4302. It allows us to extract internal outputs of a module in a simpler way and it should be a drop-in replacement of the old IntermediateLayerGetter class.

Unfortunately the create_feature_extractor() util is not currently used anywhere on vision. We should investigate using it to replace some of the old approaches. Here is a list of candidates:

• Replace IntermediateLayerGetter calls in Segmentation models [1, 2]
• Replace IntermediateLayerGetter calls in Detection models [3]
• Rewrite model surgery from the SSDFeatureExtractorVGG [4] and SSDLiteFeatureExtractorMobileNet modules [5].

Extra care needs to be given to ensure that the old weights of pre-trained models can still load and the validation statistics remain the same.

Motivation, pitch

Using the latest FX-based utility will help us future proof our code-base.

cc @datumbox

Start by reading the create_feature_extractor utility and the listed IntermediateLayerGetter uses in segmentation.py and detection/backbone_utils.py. Then inspect SSDFeatureExtractorVGG in ssd.py and SSDLiteFeatureExtractorMobileNet in ssdlite.py. Done means the candidates use the FX-based utility while old pretrained weights still load and validation statistics remain unchanged.