Add GradCache loss

[NohTow]

The GradCache method allows to scale the effective batch size of contrastive loss with a constant requirement in memory, overcoming the issue of gradient accumulation not being equivalent to larger bs for contrastive loss.
I think it would be a nice addition for people training ColBERT models with contrastive and not distillation.

As it is already implemented in Sentence Transformer, adding it to PyLate should be straightforward.
I think you already experimented with it @raphaelsty, so maybe you can take this one?

cc @tomaarsen

[raphaelsty]

I'd be happy to work on this one 🙌