Reenable SimdOps.assembleAndSum; implement Panama/Native equivalent for CosineDecoder acceleration

[michaeljmarshall]

Fixes: #367

[jkni]

This looks reasonable to me in implementing what I described in the linked issue. Basic performance testing shows the anticipated improvements for Panama/Native cosine and Panama dot product.

@marianotepper can you also take a review pass here and see if it makes sense to you? Interested in your input and also looking for a second review since I did contribute some commits to this PR.

[marianotepper]

It looks good to me.

My only suggestion would be to use a more descriptive name than decodedCosineSimilarity. Given that we are light on documentation, I would at least use a name like pqDecodedCosineSimilarity to give it some context.

[michaeljmarshall]

@marianotepper thanks for your review. I prefixed (and fixed casing) on all decodedCosineSimilarity and decoded_cosine_similarity method names with pq and pq_ respectively.

[jkni]

Agree with the name change.

[marianotepper]

LGTM