Add top-down (divisive) strategy for `hierarchical_topics()`

Feature request

hierarchical_topics() builds hierarchies exclusively bottom-up using scipy's agglomerative linkage on c-TF-IDF vectors. Add a top-down (divisive) alternative:

# Current behavior (default)
hierarchy = topic_model.hierarchical_topics(docs, strategy="agglomerative")

# New: top-down recursive splitting
hierarchy = topic_model.hierarchical_topics(docs, strategy="divisive")

Motivation

Agglomerative linkage has known limitations:

• Merge quality degrades at higher levels — late merges combine dissimilar topics because linkage minimizes global distance, not local semantic coherence
• No per-split representation — parent nodes get concatenated keywords
• #1907 — confirmed bug when 3+ topics have identical c-TF-IDF distances, breaking the hierarchy

A top-down approach avoids these issues by recursively splitting topics where each split is locally optimal.

Your contribution

I can submit a PR that adds a strategy parameter to hierarchical_topics(). The divisive path recursively splits topics using c-TF-IDF weighted NMF decomposition, building a tree where each parent-child relationship reflects a meaningful topic subdivision.

Default is "agglomerative" — existing behavior unchanged.

I've already been prototyping this in my fork (working implementation with tests). Since the divisive path is a new algorithm to maintain, I'm happy to discuss scope (e.g. landing it as experimental first) before opening the PR.

Start at the hierarchical_topics() entry point and read how the current agglomerative strategy builds the hierarchy from c-TF-IDF vectors. Review the proposed divisive path and its existing prototype tests. Done means supporting strategy="divisive" while preserving agglomerative as the default and covering the new behavior with tests.