Probabilistic backend for relational_fusion

`relational_fusion` is currently implemented using a deterministic iterative scoring approach.

I am interested in exploring whether a probabilistic method can achieve significant speedups.

Start by reading the existing relational_fusion implementation and measuring its deterministic iterative scoring approach. Define and evaluate a probabilistic alternative, with completion requiring evidence that it achieves a significant speedup without losing the intended result quality.

Written by the indexing model from the issue text.