The Bigger the Better? Rethinking the Effective Model Scale in Long-term Time Series Forecasting