Regression return logits

[noahho]

Motivation and Context

This change introduces two major improvements. First, it directly responds to the request in Issue #374 for raw logits from TabPFNClassifier. Explaining model outputs (e.g., with Shapley values) is more numerically stable and often more intuitive on an additive scale like logits. This PR adds a predict_logits method to facilitate these use cases.

Second, this work includes a significant refactoring of the internal forward and prediction methods for both TabPFNClassifier and TabPFNRegressor. This simplifies the forward pass interface, especially for the regressor, and provides a key benefit: a substantial reduction in memory usage for TabPFNRegressor. By processing ensemble outputs sequentially instead of stacking them, the regressor can now handle larger datasets and a higher number of estimators more efficiently.

---

Public API Changes

• [ ] No Public API changes
• [x] Yes, Public API changes (Details below)

Details of Public API Changes:

• TabPFNClassifier.predict_logits(X: XType) -> np.ndarray: A new public method that returns the raw, unnormalized logits for the input samples X.

• TabPFNRegressor.forward(...) Signature Change (Breaking): The forward method for the regressor has been simplified. It now returns only a single tensor of logits, instead of the previous tuple (averaged_logits, outputs, borders). This is a breaking change for users who call forward() directly (e.g., in finetuning scripts). The new interface is simpler and more memory-efficient.

• Internal Refactoring: The predict and predict_proba methods in both classes have been refactored to use the new internal logic. Their public signatures remain unchanged, ensuring backward compatibility for standard prediction workflows.

---

How Has This Been Tested?

This PR includes comprehensive new and updated test cases to ensure the correctness and consistency of the new functionality and refactored code paths:

• predict_logits Consistency: A new test, test_predict_logits_and_consistency, has been added to tests/test_classifier_interface.py. This highly parametrized test verifies that predict_logits returns output of the correct shape and type. It crucially asserts that applying softmax to the raw logits yields results that are numerically close to the probabilities from predict_proba, ensuring consistency across various configurations.

• Regressor Forward Pass: A new test test_forward_predict_logit_consistency was added to tests/test_regressor_interface.py to validate that the new, memory-efficient forward pass produces logits identical to the high-level predict method, ensuring correctness after the refactoring.

• Behavioral Tests: New tests like test_softmax_temperature_impact_on_logits_magnitude and test_balance_probabilities_alters_proba_output confirm that the modular post-processing pipeline in the classifier behaves as expected.

• Finetuning: Finetuning tests for the regressor (e.g., in tests/test_finetuning_regressor.py and examples/finetune_regressor.py) have been updated to align with the new, simplified forward method signature.

---

Checklist

• [x] The changes have been tested locally.
• [ ] Documentation has been updated (if the public API or usage changes).
• [x] An entry has been added to CHANGELOG.md (if relevant for users).
• [x] The code follows the project's style guidelines.
• [x] I have considered the impact of these changes on the public API.

[noahho]

@gemini-code-assist review