address issue for mmr calculation (#99)

I added a fix based on an issue that was addressed on the main
[langchain-ai:
max_marginal_relevance_search](https://github.com/langchain-ai/langchain/pull/22271/files)
code.

Author: IanQS

langchain_postgres/_utils.py

@@ -30,10 +30,8 @@ def cosine_similarity(X: Matrix, Y: Matrix) -> np.ndarray:
 
         X = np.array(X, dtype=np.float32)
         Y = np.array(Y, dtype=np.float32)
-        Z = 1 - simd.cdist(X, Y, metric="cosine")
-        if isinstance(Z, float):
-            return np.array([Z])
-        return np.array(Z)
+        Z = 1 - np.array(simd.cdist(X, Y, metric="cosine"))
+        return Z
     except ImportError:
         logger.debug(
             "Unable to import simsimd, defaulting to NumPy implementation. If you want "