Why Deep Models Often cannot Beat Non-deep Counterparts on Molecular Property Prediction?