Returns a Tune.KNN program whose pool is the training set.
Compilation makes no model calls. Each later run retrieves the k nearest
training rows for its input and bootstraps demonstrations from them with
the teacher role, so running the result needs a teacher configuration and
an explicit :max_model_calls, and every teacher call is counted.
Options: :k (default 3), :max_bootstrapped_demos (default k),
:max_labeled_demos (default 0), :vectorizer (function, default bag of
words).
Differences from DSPy: no embedding service is called; retrieval uses the supplied pure vectorizer. Bootstrapping accepts every successful neighbor trajectory, as DSPy does when no metric is given.