Generalizable Fault Diagnosis Under Distribution Shifts Induced by Unseen Working Conditions via Synthetic and Adversarial Sample Learning