sktime/pytorch-forecasting ? reverse-engineered prompt

Build me a Python package for time series forecasting with PyTorch.

I want it to make it easy to train modern forecasting models on pandas dataframes, especially for real world datasets with missing values, different history lengths, and multiple series IDs. It should include a simple dataset builder, a few ready to use models like Temporal Fusion Transformer, N BEATS, N HiTS, DeepAR, PatchTST, plus basic LSTM and GRU baselines. It should support multi horizon predictions, uncertainty estimates, and useful metrics, and it should have a clean training flow with PyTorch Lightning so it can run on CPU or GPU.

Please include examples, good defaults for beginners, and a way to customize or add new models later. Add logging and simple plots for actual vs predicted values. If you need current library details or best practices, look up the latest docs online first.