Avoid negative distances when using E4M3 cosine indices by renormalizing.

[psobot]

No description provided.

[dylanrb123]

@psobot I fixed the java tests and included a version bump/doc gen commit. LMK if you want me to do the latter in a separate PR.

[psobot]

Looks great, thanks @dylanrb123!

Just to talk through the implications of this change:

• Existing indices could have returned zero distances during queries due to precision errors when computing inner product
• With this change, querying existing Cosine E4M3 indices with new vectors can now return negative distances, as the normalized vectors within them will not be re-normalized. Distances will be slightly increased as the query vector is properly normalized, however.
• Newly-created indices after this change lands will never return negative distances.

One last thing I'm worried about: this change also potentially introduces quality loss when fetching a vector from the index and then passing it back in, as the vector will be re-normalized and normalize<E4M3>(x) is not idempotent. That might be something we want to think through carefully before landing this change.

[psobot]

Closing in favour of #80 due to reduced recall of this approach.