Build an open-source framework that enables training and inferencing for latent state reasoning models using Mamba architectures. This addresses the high VRAM overhead of traditional Chain-of-Thought approaches by keeping 'reasoning' in the hidden state.