VideoMAP: Toward Scalable Mamba-based Video Autoregressive Pretraining