feat: add multi-target inference

[giovp]

I've put together this PR to support 2D y arrays in TabICLClassifier and TabICLRegressor for multi-target prediction. Each target is processed independently through the existing single-target pipeline, and results are stacked.

It is useful in some contexts where I can max out GPU memory and do predictions over multiple columns in parallel. I understand the use case might be niche and so it might not be very useful for the project repo.

Thanks again for sharing weights and code!

[jonaslandsgesell]

Looks interesting!

If you are interested in building true joint multi-target regression (also for pre training) feel free to have a look into https://arxiv.org/abs/2601.05683 which describes which family of losses and architectures would be suitable.

Given the joint distribution does not factorize it should help generalization.

[JingangQu]

Hi @giovp,

Thanks a lot for this. You're right that the efficiency here is genuinely different from wrapping TabICLClassifier in sklearn's MultiOutputClassifier by keeping a single model resident on the GPU and reusing the same ensemble-generated X transform across all targets. That said, I'd like to keep TabICLClassifier and TabICLRegressor themselves single-target for now to easy the maintenance. But we could implement separate estimator, eg, TabICLMultiOutputClassifier / TabICLMultiOutputRegressor in its own module.