range_search doesn't support a radius vector

It caught my attention that range_search doesn't support a vector of radius values as an input, such that each value correspond to the radius to query each row. Would this be faster than query_radius from sklearn's KDTree? If not, is this the reason this is not implemented?

This could be a very good feature since it is the last thing needed for a faster (and accurate) calculation of Kraskov's mutual information.

Start by tracing the existing range_search API and its handling of query rows. Define how a radius vector maps to those rows, then compare the behavior and performance with sklearn's KDTree query_radius as raised in the issue. Done means the API supports per-row radii with accurate results and has coverage for the new behavior.

Written by the indexing model from the issue text.