feat: add top-down (divisive) strategy for hierarchical_topics()

[pidefrem]

What does this PR do?

feat: add top-down (divisive) strategy for hierarchical_topics()

hierarchical_topics() builds hierarchies exclusively bottom-up using scipy's agglomerative linkage. While this works for small topic counts, merge quality degrades at higher levels and it has a known bug when 3+ topics have identical c-TF-IDF distances (#1907).

Changes:

Add a strategy parameter to hierarchical_topics():

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

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

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. Each split is locally optimal, avoiding the degradation of late merges in agglomerative linkage.

Default is "agglomerative" — existing behavior unchanged.

Fixes #2508

Coordination note: this adds a new strategy option rather than changing the default, so existing hierarchies are untouched. It introduces a top-down NMF code path, which is a meaningful addition to maintain — if you'd like to discuss scope (e.g. land it as experimental first) before reviewing, I'm happy to. It also pairs naturally with the parent-node labeling in #2503.

Before submitting

• This PR fixes a typo or improves the docs (if yes, ignore all other checks!).
• Did you read the contributor guideline?
• Was this discussed/approved via a Github issue? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes (if applicable)?
• Did you write any new necessary tests?