Mitigating Overthinking in Large Reasoning Models via Manifold Steering