Add a contribution-weighted correlation plot

[guillaume-vignal]

Fixes #753

Summary

This PR adds a new correlation heatmap based on feature contributions, alongside the existing feature-correlation plot.

The goal is to visualize relationships between contributions rather than between raw input variables, making it easier to inspect how explanation signals behave across features.

How the matrix is computed

For each pair of contribution columns, Shapash computes a weighted correlation on the aligned samples:

• missing contribution values are treated as 0
• each sample is weighted by the maximum absolute contribution of the two features for that row
• the result is a symmetric correlation matrix, with the diagonal equal to 1
• values are clipped to stay in the [-1, 1] interval

In practice, this gives more importance to rows where at least one of the two contributions is strong, while keeping the same visual and clustering logic as correlations_plot.