Autoencoder-Driven Latent Representation Learning for Language-Agnostic Disordered Speech Classification Using a Universal Feature Set