`_extract_representative_docs` recomputes from scratch on every call — no caching

Nobody has claimed this yet.

### Feature request

`_extract_representative_docs()` runs the full computation pipeline (sampling → c-TF-IDF → cosine similarity → MMR selection) every time it is invoked. In interactive workflows, users often call a sequence of visualization methods that each trigger this computation:

```python
topic_model.visualize_topics() # triggers representative docs computation
topic_model.visualize_hierarchy() # triggers it again
topic_model.visualize_documents() # triggers it again
```

Cache `representative_docs_` after first computation and invalidate when topic assignments change, following the same pattern already used for `topic_embeddings_`.

### Motivation

Each call repeats the same expensive similarity computations even though topic assignments haven't changed. This is also the root cause behind [#2367](https://github.com/MaartenGr/BERTopic/issues/2367) — KeyBERTInspired recomputes representative docs embeddings even when precomputed embeddings are available.

BERTopic already caches `topic_embeddings_` with compute-once-store-invalidate-on-change semantics. Representative docs should follow the same pattern.

### Your contribution

I can submit a PR that caches `representative_docs_` after first computation and invalidates the cache when topic assignments change (i.e., after `update_topics()`, `reduce_outliers()`, `merge_topics()`). Subsequent calls return the cached result without recomputation.

**Note:** This is most effective when the underlying `_extract_representative_docs` sampling and indexing bugs are fixed first (#2495). The caching mechanism itself is independent, but caching correct results is obviously preferable.

I've already been prototyping this in my fork, so I can open a PR quickly if the approach looks good to you.

---

Start by locating _extract_representative_docs() and the existing topic_embeddings_ compute-once-store-invalidate pattern. Trace update_topics(), reduce_outliers(), and merge_topics() to identify where topic assignments change, then verify repeated visualization calls reuse the cached result and those mutations invalidate it. The issue notes that a contributor is already prototyping the change and that #2495 affects the correctness of the cached result.

Written by the indexing model from the issue text.