Fix semantic_search crashing on a list of numpy embeddings

[MohammadHijjawi97]

util.semantic_search() already accepts a 2D numpy array, but a list of 1D numpy vectors crashed in torch.stack. That is the shape SentenceTransformer.encode returns for a single string, so this public pattern failed:

query = model.encode("a query")
corpus = [model.encode(doc) for doc in documents]
semantic_search(query, corpus)

The list path now uses _convert_to_tensor, which already stacks lists of numpy arrays. A 1D corpus embedding is unsqueezed the same way a 1D query already was.

Tests: pytest tests/util/test_retrieval.py -k semantic_search
The new test fails on unmodified main and passes with this change.