Nx vs. Python performance for sentence-transformer encoding

Btw, folks can also try running ONNX versions of the models if the lack of dynamic shapes is being a hindrance.