Add crps losses to regressor

[bejaeger]

This adds Ranked Probability Score (RPS) and Ranked Logarithmic Score (RLS) to the regressor finetuning code.

Note that this PR builds upon the original work by Jonas Landsgesell and Pascal Knoll in #689. It adapts their implementation to be compatible with our latest finetuning refactor and adds RLS.

Adding the RPS loss in the finetune example results in significantly better performance.

To find good defaults for the regression loss weights, a broader benchmarking would need to be done.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[Copilot]

Pull request overview

This pull request adds RPS (Ranked Probability Score) and RLS (Ranked Logarithmic Score) losses to the regressor fine-tuning functionality and changes the default loss configuration to use RPS instead of the previous MSE-based approach.

• Refactored compute_regression_loss to _compute_regression_loss with expanded loss options including CE, RPS, RLS, MSE, and MAE
• Implemented _ranked_probability_score_loss_from_bar_logits function for computing ranked probability losses on CDF-based bar distributions
• Changed default loss weights: ce_loss_weight=0.0, rps_loss_weight=1.0, mse_loss_weight=1.0 (previously MSE was 8.0 and only loss)

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
src/tabpfn/finetuning/finetuned_regressor.py	Extended loss computation with RPS/RLS support, refactored function signature and default parameters, added comprehensive NaN handling
tests/test_finetuning_regressor.py	Added three new test functions to validate RPS/RLS loss computation, NaN masking behavior, and correct loss values; renamed existing test functions with double-underscore convention
src/tabpfn/finetuning/finetuned_classifier.py	Renamed compute_classification_loss to _compute_classification_loss for consistency with regressor changes

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[noahho]

Great job with the implementation! Had some minor comments, after which LGTM!

[jonaslandsgesell]

Great work Ben, thanks a lot for the integration in the current version. Really looking forward👍