MambaVO: Deep Visual Odometry Based on Sequential Matching Refinement and Training Smoothing