Issue: Face Embedding Normalization Can Divide by Zero

FaceNet_util_normalize_embedding divides embeddings by np.linalg.norm(embedding) without guarding against zero-length vectors. When Facenet returns a near-zero tensor (e.g., because the detected crop was empty or corrupted), the norm becomes 0 and we produce NaN/Inf embeddings. These invalid values get committed to the database and break subsequent similarity comparisons.