bugfix: apply query mask to ColBERT scores

[eelcovdw]

Fix for the issue in #75 , issue includes a minimal example and shows the proposed fix

changes:

• mask the scores with the query mask before averaging, so the average is only over the query (and not the mask tokens)
• this aligns with the stanford ColBERT implementation, where they mask the query before scoring: https://github.com/stanford-futuredata/ColBERT/blob/501c29d9e0b7f7b393e36c4177ec2b141a253114/colbert/modeling/colbert.py#L91