add different encoder ft methods

[franeknowak]

currently encoder is ft through fitting BCE onto hard labels from the same dataset. The task is so similar the encoder is likely to overfit for any task in the future. Consider adding image reconstruction, other methods for encoder pretraining.

[franeknowak]

do we actually want to fine-tune the encoder risking the overfitting. IMNT weights show good results - review in the future.