fix: return empty list for non-positive top_k in cosine_topk

[koriyoshi2041]

cosine_topk returns the entire corpus when top_k is 0, and n-1 results when top_k is negative, instead of returning nothing.

src/memu/database/inmemory/vector.py:

actual_k = min(k, n)
if actual_k == n:
    topk_indices = np.argsort(scores)[::-1]
else:
    topk_indices = np.argpartition(scores, -actual_k)[-actual_k:]   # [-0:] == [0:] -> all
    ...

For k == 0, actual_k is 0, and the else branch slices [-0:] which is [0:] — the whole array. For k == -1 it slices [1:] (n-1 items).

top_k reaches cosine_topk unvalidated from retrieve_config.{category,item,resource}.top_k, so a user setting a tier's top_k to 0 to disable it silently gets the full corpus back rather than an empty result — an over-retrieval bug that also inflates token usage downstream.

Fix: return [] for k <= 0. cosine_topk_salience is guarded the same way — scored[:0] already returned [], but scored[:-1] (for k = -1) incorrectly dropped the last entry, so both now behave consistently.

Added tests/test_vector.py covering k=0/k=-1 (empty) and a positive-k ordering case for both functions. pytest tests/test_vector.py passes; ruff clean.