New Index Predicate types using Embeddings

[fgregg]

it would be really fun to try a new index-type predicate that used a pre-trained embedding model like Universal Sentence Encoder and approximate nearest neighbor, i.e. annoy.

Maybe this might be of interest to you, @lmores.

This should probably be an optional features, since large language models are so large.