XGRegressor::fit panics on a zero-row training set

[teddytennant]

XGRegressor::fit panics on a training set with zero rows, with default parameters and no subsample set. It hits the same line as #444 but the cause is different, so #445 does not fix it. I confirmed this panic still happens with #445 applied.

A zero-row matrix is reachable through the public Array2::take:

use smartcore::linalg::basic::matrix::DenseMatrix;
use smartcore::linalg::basic::arrays::{Array, Array2};
use smartcore::xgboost::{XGRegressor, XGRegressorParameters};

let full = DenseMatrix::from_2d_vec(&vec![vec![1.0, 1.0], vec![2.0, 1.0]]).unwrap();
let empty = full.take(&[] as &[usize], 0);
println!("shape = {:?}", empty.shape());   // (0, 2)

let y: Vec<f64> = vec![];
let _ = XGRegressor::fit(&empty, &y, XGRegressorParameters::default());

shape = (0, 2)
thread 'main' panicked at src/xgboost/xgb_regressor.rs:323:21:
attempt to subtract with overflow

find_best_split runs 0..sorted_idxs.len() - 1, which underflows on an empty slice whatever put it there. In release builds, where overflow checks are off, this comes out as index out of bounds: the len is 0 but the index is 0 instead.

The right behaviour is a design call, which is why I have not sent a patch. Either fit returns Failed::because(FailedError::ParametersError, ...) for empty training data, which matches the direction of #435, or it returns Ok with a degenerate model. I lean towards the error, since a model trained on no data is not useful, but it is your call and I am happy to send whichever you prefer.