Truncate recommend() result to caller's N when filter_items is set (#736)

[Cyberfilo]

Summary

`ItemItemRecommender.recommend(N=K, filter_items=F)` was returning up to `K + len(F)` items instead of `K`. Fixes #736.

Root cause

In `implicit/nearest_neighbours.py::ItemItemRecommender.recommend`, the post-mask slice reused an inflated `N` that was meant only for the over-fetch step:

```python
if filter_items is not None:
N += len(filter_items)
...
if filter_items is not None:
mask = np.isin(ids, filter_items, invert=True)
ids, scores = ids[mask][:N], scores[mask][:N] # uses inflated N
```

Fix

Store the caller-requested `N` before inflating, then truncate to that value after masking:

```python
requested_n = N
if filter_items is not None:
N += len(filter_items)
...
if filter_items is not None:
mask = np.isin(ids, filter_items, invert=True)
ids, scores = ids[mask][:requested_n], scores[mask][:requested_n]
```

The over-fetch is still required to ensure enough rows survive the mask in the worst case (when every filtered item happens to be in the top-N scoring window). No behavior change for callers that don't use `filter_items`.

Test plan

Manual reproducer:

```python
from implicit.nearest_neighbours import CosineRecommender
from scipy.sparse import csr_matrix
import numpy as np

model = CosineRecommender(K=20)
model.fit(csr_matrix(np.random.rand(10, 10)))
user_items = csr_matrix(np.ones((1, 10)))
ids, _ = model.recommend(0, user_items, N=3, filter_items=[1, 2, 3], filter_already_liked_items=False)
assert len(ids) == 3 # before: len(ids) == 6
```

Diff is 5 lines: 1 variable assignment, 1 comment block, and a 1-character change on each of two slice lines.