Minor change to TabM seach space?

[dholzmueller]

The current TabM search space tunes n_layers in [2, 5] but the original search space with numerical embeddings tunes [1, 4]: https://github.com/yandex-research/tabm/blob/main/exp/tabm-mini-piecewiselinear/adult/0-tuning.toml
In correspondence with Yury Gorishniy we agreed to increase the lower limit to 2 but I think an upper limit of 4 should still be "more correct".

[LennartPurucker]

Note, we can connect this to re-running TabM once we have its new implementation ready.

[LennartPurucker]

The new implementation is now public: https://github.com/yandex-research/tabm#advanced-usage

Thus, we need to re-implement our version of TabM.

[LennartPurucker]

Current code branch: https://github.com/autogluon/tabrepo/tree/prep_for_ag140_tabm

Blocking on yandex-research/tabm#13

[LennartPurucker]

TODO: add new onehot logic

[dholzmueller]

Gradient clipping is not effective in the current implementation, it needs to be fixed (depending on whether it should actually be used with the corrected implementation, since I don't know if it was actually used in the original TabM results).
yandex-research/tabm#20
dholzmueller/pytabkit#34