On the Efficacy of Differentially Private Few-shot Image Classification