ngushchin/LightSB ? reverse-engineered prompt

Build me this Light Schrödinger Bridge project so I can reproduce the main experiments from the paper. I want the code organized around the notebooks for the toy examples, swiss roll, benchmark, single cell, and image translation with ALAE, with the core LightSB logic living in reusable Python modules. Please make it easy to run the experiments end to end, load the example data, generate the plots and teaser style results, and save outputs in a sensible way. If anything depends on current PyTorch or notebook tooling details, look up the latest docs online if you need to. I also want the training and experiment tracking flow to work with Weights and Biases, since that is part of how this repo is meant to be used. Keep it practical and reproducible, with clear notebook steps and code that matches the paper as closely as possible.