Improve interactions plot parity with contribution plot

[guillaume-vignal]

interactions_plot and top_interactions_plot were not fully aligned with contribution_plot for multiclass handling and interaction readability.

Description

• Added jittered violin point dispersion for interaction plots (with true non-jittered hover values).
• Added adaptive interaction color scaling (cmin / cmax).
• Added robust x-axis categorical mapping persistence in top interactions dropdown updates.
• Added pair auto-ordering rules for readability (auto_order=True by default).
• Added low-cardinality numeric-as-categorical handling on x-axis.
• Switched interaction display values to display-ready data (x_init) for postprocessing/transcoding consistency.
• Added crossed smart sampling on the joint distribution of the two interaction features.
• Added multiclass label support:
  • interactions_plot(..., label=...)
  • top_interactions_plot(..., label=...)
  • propagated to report interaction generation.
• Hardened SHAP interaction output handling for multiclass/multi-output formats (list and 4D variants).
• Added type hints in plot_interactions.py.

Closes #757.