MARS6: A Small and Robust Hierarchical-Codec Text-to-Speech Model