Server-Side Local Gradient Averaging and Learning Rate Acceleration for Scalable Split Learning