Towards Robust Assessment of Pathological Voices via Combined Low-Level Descriptors and Foundation Model Representations