Activation Steering in Neural Theorem Provers