PriorLabs/TabPFN ? reverse-engineered prompt

Build me a Python library for fast tabular machine learning that can do classification and regression with a simple fit and predict workflow.

I want it to feel easy to use for someone with a pandas DataFrame or numpy arrays, where they can train on their tabular data and get predictions right away. Please include a clean public API for a classifier and regressor, support a few model versions including a faster one, and make it download the model weights on first use. It should work well on modern Python versions, run locally with PyTorch, and prefer GPU when available but still work on CPU for smaller datasets.

Please also add a few example scripts and a short notebook style demo so someone can quickly try binary classification, multiclass classification, and regression. If there are any setup details or current best practices from the docs that would help, look them up online and follow them.