Query Drift Compensation: Enabling Compatibility in Continual Learning of Retrieval Embedding Models