[classify.scripts.saliency] Incorporate evaluation code directly on interpretability/completeness analysis; Remove outdated saliency-evaluation script and engine; Fix test units

[classify.engine.saliency.interpretability] Implement filter for operating on specific datasets only; Implement basic analysis for interpretability

[libs.common.models.transforms] Maximise use of torchvision transforms; Closes #86 as well concluding previous commit

[libs.segmentation.config.models.lwnet] Set default cosine annealing strategy to be closer to original publication

[libs.common.engine.callbacks] Use appropriate interface for extracting the current learning rate from schedulers

[evaluation] Homogenize segmentation/classification code; Use more credible constructs (DRY); Fix test units

[*/models] Unify modelling by applying DRY; Fix a nasty loss-balancing bug; Make loss-configuration private to base model class; Automate loss-configuration; Fix typing across model submodules; Move `name` into parent type