Adapt the MMR computations to the actual metric being used in the vstore

[hemidactylus]

Ah, ... there's a more conceptual thing that's bugging me (and to be clear it also affects the VectorStore I myself wrote, lol).
We offer all 3 similarities Astra has (ok, dot == cosine where it makes sense to use it. But still there's Euclidean).
Yet the MMR always assumes a cosine metric to do its MMR thing. I think in some cases it could lead to very suboptimal/surprising reranking results.
I would consider filing an issue (for the future) whereby the measure used within the MMR step matches the one of the vector store (which is probably what the user intuitively expects).
(granted, lambda_mult values would not have the same meaning across metrics, ... so what? Folks would tune to their metric of choice)