Bias Mitigated Learning from Differentially Private Synthetic Data: A Cautionary Tale