Clear all cached tensor references. Call between training batches to prevent memory accumulation from intermediate tensors. Tensors will be re-registered on the next forward pass.