Cross-encoder distillation

[raphaelsty]

Working on cross-encoder distillation

We want to distill cross-encoder into ColBERT,

• Update the training so we can distill logits pre-computed on option.
• Create a function which computes logits using a pre-trained cross-encoder in order to create a dataset with logits that can be used to distill soft labels to colbert.

[NohTow]

I think we can close this now.