Make model initialization configurable and not hard coded - pretrained=True option

[baraklior]

My problem
I work in a production environment with no internet access at runtime and i'm loading weights from a file.
when initializing a FeaturExtractor it in turn loads a model with a hard coded argument pretrained=True, so it attempts to download the model weights.
specifically i'm looking at this line but it pops up in a few places:
self.feature_extractor = FeatureExtractor(backbone=_backbone(pretrained=True), layers=["avgpool"]).eval()

proposed solution
make the pretrained option to be configurable by environment variables / as an init argument / config option in the classes -
TorchInferencer
DfkdeModel
FeatureExtractor
and any other classes that initialize a model

the offending file
https://github.com/openvinotoolkit/anomalib/blob/cab7aa21aba6876173585a6d300c63238b16fb11/anomalib/models/dfkde/torch_model.py

[samet-akcay]

@baraklior, just to understand the request correctly, you want to add pre_trained flag to config files to set it on/off depending on the need, right? For example, something like this for DFKDE: