Training on Fake Labels: Mitigating Label Leakage in Split Learning via Secure Dimension Transformation