Ternary Mamba: Grouped Quantization-Aware Training of W1.58A16 State Space Models