Change NNDescent to USearch in approximate Butina split

[j-adamczyk]

We currently use NNDescent in approximate Butina split. USearch can be used instead, which has nearly identical interface, but should be faster and also has no additional dependencies (single C++ binary + Python interface).