XTR for multivector search doesn't provide implementation details

[prrao87]

The multivector search documentation mentions that "LanceDB also integrates with ConteXtualized Token Retriever (XTR)" but provides no implementation details, code examples, or usage instructions. Users reading this see XTR exists but have no idea how to use it in LanceDB. Some ideas below.

The XTR (ConteXtualized Token Retriever) model for multi-vector search can be accessed primarily through open-source implementations and Hugging Face. It is designed to enhance search quality by prioritizing the most semantically important document tokens, acting as a faster alternative to ColBERT.