enable ONNX export

[AlexanderPfefferle]

modified the code to support ONNX export:

• use custom implementation of torch.nanmean and torch.nansum everywhere
• replaced usage of torch.Generator with code that isolates the rng in add_embeddings

[LeoGrin]

LGTM, thanks a lot @AlexanderPfefferle 🥳🥳 I also feel like it makes the code simpler, so that's great!
The pre-commit error is not related to this PR, I've just fixed it on main.
For the regressor test, IIRC, the change is because check_methods_subset_invariance now also fail close to the 1e-7 threshold for lack of precision, correct?

[LeoGrin]

Thinking about it, would it be easy to add a test that TabPFN is ONNX exportable, to make it sure it stays exportable?

[AlexanderPfefferle]

Yes exactly, check_methods_subset_invariance is failing on the main branch right now with Max absolute difference among violations: 3.199093e-06.
I've added tests that check for exportability on CPU.

[LeoGrin]

Thanks a lot @AlexanderPfefferle ! onnx should be added to the dependencies (but not the main ones) to make the tests green. (I can also do it if you add me to your fork).

[LeoGrin]

I set the random state carefully to pass the regressor test for older versions, which is due to #175 and not to this pr.

[LeoGrin]

Merging, thanks a lot @AlexanderPfefferle !