BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models