Batch processing is causing inconsistencies

[roquelopez]

When we run the ct_learning method (contrastive learning), it produces inconsistent matches for the same column under different scenarios. For example, if we provide the source columns A, B, and C together, the method returns the pair (A, Target1). However, if we provide only the column A as input, it returns (A, Target2).

Upon forcing batch_size=1 during inference time, the results become consistent. This behavior suggests a potential issue with the padding strategy or inference in the batches.

[lyrain2001]

Hi Roque, I believe you are right. I didn't fully grasp the problem when Eden discussed it with me yesterday, but I now understand that the issue is related to padding. I created a completely new version for Magneto, so this issue slipped my mind. We now have two options:

1. Update the entire bdi-kit codebase to incorporate Magneto's version.
2. Rewrite this particular part.
   Could you please let me know which option you prefer? Thank you very much!