CV Train Classifier

Trains a cv2.ml classifier and evaluates it, all inside one node: the features are split into a train and a held-out test set (stratified per class, deterministic per seed), the picked algorithm is fitted on the train split and scored on the test split (accuracy + confusion matrix). The model never leaves the node - train, evaluate and the optional query prediction all happen in one call, so caching stays correct. Connect query_features (e.g. an 'CV Coordinate Grid') to also classify arbitrary extra samples with the freshly trained model - reshape + colormap those predictions to render decision regions. Degenerate data (no samples, a single class, an algorithm that cannot fit - e.g. Boost with 3 classes) yields success=false with safe empty outputs and query predictions of the first class, never an error. model_yaml carries the serialized model for reuse outside ComfyUI (cv2.ml.*_load).