DITTO: Data-efficient and Fair Targeted Subset Selection for ASR Accent Adaptation