model_architecture.py updates

[Raindrop182]

• For the ConvNeXtBase architecture, the tune argument should be False by default, instead of True, to match the behavior of the other architectures and to ensure frozen epochs during training behave as intended.
  • Frozen Epoch Bug: train.py loads in models using load_classifier which doesn't explicit set the tune argument to False. Since ConvNeXt is the only model with tune=True as the default, sometimes ConvNeXt would be loaded in with all parameters unfrozen even during the beginning frozen epoch stage.
• The forward function for EfficientNet can be simplified to return self.model(x). It yields identical results and makes the code cleaner.