MoMa: Efficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts