Implement self-organizing maps (SOMs) in `linfa-clustering`

If possible, I'd like to add an implementation of self-organizing maps to linfa-clustering. Unlike many of the clustering algorithms already implemented in linfa, SOMs preserve topological relationships, offering a continuous map from n-dimensional space to a 2D grid. This makes SOMs particularly useful for visualizing higher-dimensional data compared to many alternatives such as k-means clustering, which simply partitions the data into clusters without necessarily maintaining spatial relationships or "distance" between points.

Start by reviewing the existing clustering algorithms in linfa-clustering, especially the k-means implementation, and compare their APIs with the requirements of self-organizing maps. Clarify the expected 2D grid behavior, training process, and visualization-oriented output before implementation; the issue currently does not define concrete acceptance criteria for completion.