Improve confusion matrix readability on imbalanced data

[ZakariaRida96]

Description
This PR improves confusion matrix heatmap readability for highly imbalanced datasets (see #685).

Previously, the color scale was based on the maximum value, making smaller values hard to distinguish when one class dominates.
This update uses a quantile-based upper bound and clips extreme values to enhance contrast while keeping the visualization clear.

Main changes:

• Add quantile-based scaling to plot_confusion_matrix
• Add optional parameters plot_confusion_matrix and also through SmartPlotter.confusion_matrix_plot
  • quantile=0.95 by default
  • use_quantile_scale=True by default
• Clip extreme confusion matrix values before rendering the heatmap, depending of the chosen quantile threshold
• Keep backward compatibility by allowing the feature to be disabled
• Add unit tests for quantile scaling and disabled scaling

Actuel behavior:

• Large dominant values no longer flatten the full color scale
• Smaller values become more visible in the plot
• The rendered matrix is clipped using the chosen quantile threshold
• Display both real and clipped counts in hover text when the scale method is actived
• The feature can be disabled to restore the previous full-range behavior

Here is the application on the exemple shown on the issue:

# Use the full value range for the color scale
plot_confusion_matrix(
    y_true=y_true_labels,
    y_pred=y_pred_labels,
    width=500,
    height=400,
    use_quantile_scale=False,
)

---

# Use the 90th percentile as upper bound for the color scale
    y_true=y_true_labels,
    y_pred=y_pred_labels,
    width=500,
    height=400,
    quantile=0.90,
)

---