M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models