`approximate_distribution`: add target-based min_similarity ("at most N unassigned documents")

Nobody has claimed this yet.

### Feature request

`approximate_distribution()` requires manually specifying `min_similarity`. Like the `threshold` parameter in `reduce_outliers`, users must iterate to find a value that produces an acceptable number of documents with zero topic distribution.

Add an `outliers_nb_target` parameter that auto-searches for the optimal `min_similarity`:

```python
# Current: manual trial-and-error
distributions = topic_model.approximate_distribution(docs, min_similarity=0.1)
# Count zeros... too many. Try 0.05... too few. Try 0.08...

# Proposed: specify the target directly
distributions = topic_model.approximate_distribution(docs, outliers_nb_target=50)
```

### Motivation

Users know their target — "I want at most 50 documents without any topic assignment" — not the internal `min_similarity` value that achieves it. The same trial-and-error UX problem as `reduce_outliers` (see related issue).

### Your contribution

I can submit a PR that adds:
- `outliers_nb_target` (int): target number of zero-distribution documents
- **Performance optimization:** refactor the internals so that the expensive similarity matrix is computed once and reused across all binary search iterations

Cannot set both `min_similarity` and `outliers_nb_target`. Backward compatible: defaults to `None`.

I've already been prototyping this in my fork, so I can open a PR quickly if this looks like a good direction.

---

Start with the approximate_distribution() implementation and compare its threshold-search behavior with reduce_outliers. Define how outliers_nb_target interacts with min_similarity, then verify that the similarity matrix is reused across the search and that the default None behavior remains backward compatible.

Written by the indexing model from the issue text.