Use MTLR to fit logistic regressions in DebiasedBCE

[GabrielGimenez]

Out current bottleneck in DebiasedBCE training time is logistic regression training.
We can try to replace scikit learn logistic regression with a custom MTLR module.

The most promising is to use a multi-task logistic regression approach, so we maintain multiple outputs but fit only one model. I think the most straightforward way to do this is via torch, which opens up the possibility of unlocking GPUs across the board. The drawback is to have torch as a dependency to the lib.