Add TabpfnEstimator

[jchung02]

This PR adds TabpfnEstimator, bringing TabPFN (Prior-Data Fitted Networks) into DoWhy's estimator suite as an outcome model for backdoor adjustment.

TabPFN is a tabular foundation model pre-trained on synthetic data — it requires no hyperparameter tuning and works out of the box, making it a practical option for small-to-medium tabular datasets (≤10,000 samples, ≤500 features).

Key Features

• Automatic model type detection: selects TabPFNClassifier or TabPFNRegressor based on outcome dtype and cardinality
• Single-GPU and multi-GPU inference support; multi-GPU mode distributes n_estimators across devices.
• Bootstrap-based confidence interval and significance testing
• Informative warnings for TabPFN pretraining limit violations (sample size, feature count, class count)

Added Files

• dowhy/causal_estimators/tabpfn_estimator.py
• tests/causal_estimators/test_tabpfn_estimator.py
• docs/source/example_notebooks/dowhy_tabpfn_estimator.ipynb

Modified Files

• docs/source/example_notebooks/dowhy_estimation_methods.ipynb
  — Added Method 9: TabPFN Estimator section

Dependencies

TabpfnEstimator requires additional packages not included in DoWhy's default dependencies:

pip install tabpfn torch

Installation instructions and authentication guidance for TabPFN v2.5+ are included in the source file and example notebooks.

References

1. Hollmann, N., Müller, S., Eggensperger, K., and Hutter, F.
   TabPFN: A Prior-Data Fitted Network for Tabular Data.
   ICLR (2023). https://arxiv.org/abs/2207.01848
2. Hollmann, N., Müller, S., Purucker, L., Krishnakumar, A., Körner, M.,
   Hoo, R., Schirrmeister, R. T., and Hutter, F.
   Accurate predictions on small data with a tabular foundation model.
   Nature (2025). https://doi.org/10.1038/s41586-024-08328-6