Implement Kolmogorov-Smirnov test and integrate into `ClassificationReport`

[MoonFlowww]

Motivation

The KS test is a distribution-free way to check whether a classifier's predicted probability distribution matches the empirical distribution of true labels. It is widely used as a calibration diagnostic alongside reliability diagrams, especially in binary classification. Nott already has calibration infrastructure (Temperature scaling, reliability diagrams, ROC/PR curves) — the KS statistic is the natural next addition and its import is already stubbed.

Current State

src/evaluation/details/classification.hpp:414:

using Nott::Plot::Details::compute_kolmogorov_smirnov; // TODO: KS

The function is imported (or forward-declared) from the plot module but never called, and ClassificationReport has no field for the result. compute_kolmogorov_smirnov itself appears to be unimplemented.

Proposed Change

1. Implement compute_kolmogorov_smirnov in src/plot/:

struct KSResult { double statistic; double p_value; };

KSResult compute_kolmogorov_smirnov(const torch::Tensor& predicted_probs,
                                     const torch::Tensor& true_labels);

Algorithm:

• Sort samples by predicted probability
• Compute empirical CDFs for positive and negative classes separately
• KS statistic = max |CDF_pos(t) - CDF_neg(t)| over all thresholds t
• p-value via the standard KS distribution approximation

2. Add fields to ClassificationReport:

struct ClassificationReport {
    // ... existing fields ...
    double ks_statistic = 0.0;
    double ks_p_value   = 1.0;
};

3. Call it inside the binary classification evaluation path and populate the new fields.

Acceptance Criteria

• compute_kolmogorov_smirnov implemented and returns {statistic, p_value}
• ClassificationReport exposes ks_statistic and ks_p_value
• Only computed for binary classification (skip / set NaN for multi-class)
• // TODO: KS comment removed
• Numerically consistent with scipy.stats.ks_2samp for equivalent inputs