Implement effective $\varepsilon$ as a metric.

[fhoussiau]

• Add a ROC-curve metric, $(FP_\tau, TP_\tau)_{\tau\in\mathbb{R}}$, to compare attacks.
• Add a simple estimation of $\varepsilon^\text{eff}$ as $\max_{\tau} \frac{TP_\tau}{FP_\tau}$ (with some minimum number of samples to make sure that the estimate is well-defined).
• Implement Clopper-Pearson bounds confidence intervals for TP and FP, in order to have a statistically-significant estimate of $\varepsilon^{eff}$.