[models,configs/models] Cleanup model implementation to re-use all backbones from torchvision; Simplify checkpointer; Remove own implementation of model_zoo; Implement normalization for torchvision-based backbones; Enables pytorch installation on osx

[script,engine,test] Allows the user to specify the number of steps for threshold/performance analysis

[engine.*] Move model summarization to training; Separate static information from dynamic information logging

[script.*] Implement F1-score a priori where possible; Improve plotting of precision-recall curves with better legends, more consistence; Remove redundant plotting code; Update tests

This reverts commit 4edb6d5b.