Refuse a 0-row matrix in DBSCAN::predict

[SomeB1oody]

Problem

DBSCAN::predict returns an empty array for a matrix with 0 rows. Every other model returns EmptyInput. The early return also comes before the feature-count check. A 0-row matrix with the wrong feature count therefore returns Ok, where it should return an error.

Reproduction

use ndarray::{Array2, array};
use rustyml::machine_learning::DBSCAN;

fn main() {
    let x = array![[0.0, 0.0], [0.1, 0.1], [0.2, 0.0], [5.0, 5.0], [5.1, 5.1], [5.0, 5.2]];
    let mut model = DBSCAN::new(0.5, 2).unwrap();
    model.fit(&x).unwrap();

    // The model has 2 features. Both 0-row matrices return Ok
    println!("0x2: {:?}", model.predict(&Array2::<f64>::zeros((0, 2))));
    println!("0x3: {:?}", model.predict(&Array2::<f64>::zeros((0, 3))));
}

Output on master:

0x2: Ok([], shape=[0], strides=[0], layout=CFcf (0xf), const ndim=1)
0x3: Ok([], shape=[0], strides=[0], layout=CFcf (0xf), const ndim=1)

Fix

This PR removes the early return, so predict goes through validate_predict_input like every other model. This is a behavior change, and CHANGELOG.md marks it as one. A fit that finds no core point still stores a (0, n_features) matrix, so the feature-count check stays correct for that model.

Output on this branch:

0x2: Err(EmptyInput("dataset to predict on"))
0x3: Err(EmptyInput("dataset to predict on"))