Standardising matrix arrays (adopt Array API spec)

[KennethEnevoldsen]

Currently MTEB does not consistently use a ground truth for vector manipulation. A solution would be to adopt the Array API spec. A potentially simpler solution is to use either torch or numpy.

• The benefit of using numpy is that it is very stable and rarely leads to OOM and related GPU issues.
• The benefit of using PyTorch is that it potentially allows evaluation using GPUs (currently, sklearn does not support this, but might in the future, when they adopt they API spec)

partly related to the discussion in: #883

This is not a pressing concern atm.

[Muennighoff]

Good point; I think using numpy would also allow us to make torch an optional dependency which could be nice as I assume some models do not require it so people have to install it just for mteb